Hi folks,
This summer I’m working on improving dominators during my internship at Google. Below is an RFC on switching to dynamic dominators, which you can also read as a Google Doc if you prefer so. Please let us know what you think.
~Kuba
Hi folks,
This summer I'm working on improving dominators during my internship at
Google. Below is an RFC on switching to dynamic dominators, which you can
also read as a Google Doc
<Dynamic dominators - Google Docs;
if you prefer so. Please let us know what you think.~Kuba
=======================================================================
*1. Context*
Currently LLVM uses the Simple Lengauer-Tarjan algorithm to build the
(post)dominator tree. The only way to update it is by manually setting
IDoms which is not obvious in many cases and can be extremely error-prone.
And because updating it manually is so difficult, programmers tend to just
recompute it after changing the CFG (by not AU.addPreserved()'ing the
DomTree). This causes DomTree calculation to fire very often even if only a
very small portion of it gets really affected by the changes in CFG. As an
example, when optimizing a Full LTO clang bitcode, DominatorTreeWrapperPass
alone calls DT.recalculate over 6.5 million times, which takes 20s on my
machine.
Assuming an optimistic 10min clang FullLTO build time, this is about 3%, so
overall this is actually a pretty perf improvement I think.
Using an incremental algorithm it would be much easier to keep an
up-to-date DomTree without custom update logic, which will save us the time
currently spent during DomTree recalculations and reduce the number of bugs
caused by manual updates.
Avoiding bugs seems really compelling. (IIRC talking with you, Davide, and
DannyB at the social, there are tons of such bugs, so eliminating them by
mechanically just using a new API seems awesome)
It would also make it feasible to maintain postdominators and use them
more broadly, which currently can be too complicated and expensive.*2. Proposal*
The proposal is to make dominators use the incremental algorithm that
allows to keep (post)dominator tree up to date as CFG changes. To achieve
that, we would implement the dynamic depth based search algorithm (DBS)
described in this paper [1] <https://arxiv.org/pdf/1604.02711.pdf> and
expose 2 main new functions: insertArc(From, To) and deleteArc(From, To).
The algorithm uses SemiNCA under the hood which would replace
Lengauer-Tarjan implementation currently used.The second part of the proposal is to gradually deprecate and remove the
existing API for manually manipulating dominator tree
(changeImmediateDominator, addNewBlock) and replace its use within LLVM
with the new incremental API.*3. Proof of concept*
The prototype implementation can be found in my LLVM fork [2]
<https://github.com/kuhar/llvm-dominators>\. It comes with several layers
of verification and was tested on clang, llvm test suite and a few open
source libraries.
The code provides the basic functionality and tries be ‘API-equivalent’
with the current DomTree implementation. The NewDomTree is also able to
convert itself to the current one for testing and verification purposes.
Dynamic dominators are hacked into 3 existing passes (DomTreeWrapperPass,
LoopDeletion, LoopSimplifyCFG) to test correctness and experiment with the
use of the new API.The prototype also comes with a bunch of testing utilities and a simple
benchmarking tool.*4. Performance*
The real life performance of full dynamic DBS-based DomTree recalculation
is between 20% and 2% better on a machine with two Xeon E5-2680v2 CPUs than
the existing SLT algorithm, which is consistent with the findings from this
thesis [3] <ftp://ftp.cs.princeton.edu/reports/2005/737.pdf>\. The
advantage is not that visible on debug builds, where the DBS-algorithm can
be up to 8% slower. It is most like possible to speed up debug builds, but
I haven’t looked into that yet.
The benchmark performed [4]
<Dynamic dominators - Google Docs;
loads of a (full LTO) bitcode file and builds DomTress of all functions 15
times.The current DomTree updates DFS in-out numbers eagerly upon construction,
while the new one postpones is until they are actually needed. To make the
benchmark fair, numbers were collected for both eager and lazy strategy for
the new DomTree.The main advantage of the incremental algorithm comes from the fact that
it allows incremental updates without rebuilding the whole tree, not from
the slightly faster construction.
What’s more, the insertArc / deleteArc API doesn’t force updates to happen
immediately -- they can be queued behind the scenes and happen in batches
if we decide to pursue that design later.*5. Transition plan*
There are several possibilities when it comes to transition. The biggest
problem is that the current DomTree exposes an interface for manual updates
(setIDom, changeImmediateDominator), and for manual construction
(addNewBlock). Because of the additional data stored in the incremental
algorithm (relative dominators, preorder parents, levels) it is not really
possible to use the old API while keeping this data up-to-date. The
incremental algorithm relies on this information when performing fast arc
deletions; It is still able to perform them without it -- deletions are
then just slower in some cases.
The most straightforward solutions are:a) Keep the existing DomTree and gradually replace its uses with the new
one. It is possible to convert the DBS-based dominators to the old ones.
b) Replace the existing DomTree with the new, dynamic dominators. Nuke all
of the old update functions and replace their uses with the new incremental
API in one commit.
c) Replace the existing DomTree with the new one, but keep the old API
around and mark it as deprecated. If someone invokes one of the old update
functions, mark the the additional information as invalid for dynamic
deletions. Follow the pessimistic but correct dynamic deletion path if the
additional information is marked as invalidated. Gradually modernize the
projects to use the new API instead and then remove the old update
functions.Maintaining the old DT and the new one simultaneously seems like a waste
of time as the DBS offers better or similar performance to the existing
SLT-based implementation.
The problem with replacing the old API with the new one is that it’s used
in many places (~100) across the project and I believe that doing that all
at once would be tricky to get correct. Gradual update seems much easier to
ensure correctness and the transitional API (invalid additional data check)
is trivial to implement. Because of that, I propose to follow the option
(c).
(c) sounds like the best choice to me too. That would also allow fixing
dominator verification failures first, giving a nice immediate benefit to
motivate and kickstart the transition.
-- Sean Silva
Hi folks,
This summer I'm working on improving dominators during my internship at
Google. Below is an RFC on switching to dynamic dominators, which you can
also read as a Google Doc
<Dynamic dominators - Google Docs;
if you prefer so. Please let us know what you think.~Kuba
=======================================================================
*1. Context*
Currently LLVM uses the Simple Lengauer-Tarjan algorithm to build the
(post)dominator tree. The only way to update it is by manually setting
IDoms which is not obvious in many cases and can be extremely error-prone.
And because updating it manually is so difficult, programmers tend to just
recompute it after changing the CFG (by not AU.addPreserved()'ing the
DomTree). This causes DomTree calculation to fire very often even if only a
very small portion of it gets really affected by the changes in CFG. As an
example, when optimizing a Full LTO clang bitcode, DominatorTreeWrapperPass
alone calls DT.recalculate over 6.5 million times, which takes 20s on my
machine.Assuming an optimistic 10min clang FullLTO build time, this is about 3%,
so overall this is actually a pretty perf improvement I think.
s/pretty/pretty small overall/
Hi Jakub,
thanks for the update. I was already eagerly waiting to hear what your
internship on dominance brings. I think the numbers are already very
encouraging and I believe moving incrementally over to the new algorithm
is exactly the right approach.
I have a couple of questions:
1) Daniel and Chandler have for a long time been talking about computing
dominance and post-dominance in one shot to reduce the cost of
post-dominance and make it (freely) available everywhere. Is this
covered by your current (or planned) work?
2) I tried to use tools/dominators
I wanted to play a little bit with your work, but run in some issues:
bin/llvm-as ../test/Analysis/Dominators/2007-07-11-SplitBlock.ll
bin/dominators -to-graph ../test/Analysis/Dominators/2007-07-11-SplitBlock.bc | tee graph
p 5 5 1 1
a 1 3
a 1 5
a 2 3
a 3 2
a 3 4
e
bin/dominators graph
Unknown file format for graph
Invalid input graph
I must do something obvious wrong. Any idea what?
Best,
Tobias
Hi Tobias,
- Daniel and Chandler have for a long time been talking about computing
dominance and post-dominance in one shot to reduce the cost of
post-dominance and make it (freely) available everywhere. Is this
covered by your current (or planned) work?
I’m not sure what you exactly mean by one shot; I’ll ask around tomorrow.
I wanted to play a little bit with your work, but run in some issues:
bin/llvm-as …/test/Analysis/Dominators/2007-07-11-SplitBlock.ll
bin/dominators -to-graph …/test/Analysis/Dominators/2007-07-11-SplitBlock.bc | tee graph
p 5 5 1 1
a 1 3
a 1 5
a 2 3
a 3 2
a 3 4
e
bin/dominators graph
Unknown file format for graph
Invalid input graph
I must do something obvious wrong. Any idea what?
You almost got it right – ‘graph’ files must end with “.txt”… 
Best,
Kuba
Hi Tobias,
1) Daniel and Chandler have for a long time been talking about computing
> dominance and post-dominance in one shot to reduce the cost of
> post-dominance and make it (freely) available everywhere. Is this
> covered by your current (or planned) work?I'm not sure what you exactly mean by one shot; I'll ask around tomorrow.
Not sure what algorithm Daniel had in mind, but he was talking about
some approach that basically gives post-dominance "for free" when
computing dominance. Not sure how exactly this would work.
I wanted to play a little bit with your work, but run in some issues:
> > bin/llvm-as ../test/Analysis/Dominators/2007-07-11-SplitBlock.ll
> > bin/dominators -to-graph ../test/Analysis/Dominators/2007-07-11-SplitBlock.bc
> > tee graph
> p 5 5 1 1
> a 1 3
> a 1 5
> a 2 3
> a 3 2
> a 3 4
> e
> > bin/dominators graph
> Unknown file format for graph
> Invalid input graph
> I must do something obvious wrong. Any idea what?You almost got it right -- 'graph' files must end with ".txt"...
I got it. Thank you! I also realized I can directly read in .bc files.
I have a couple of immediate (but rather minor) comments:
- It would be convenient to allow .ll files to be read.
- You do not use the BB names in the output, right? This would certainly
improve readability!
- My output prints "New Dominator Tree" to times in a row. What is the
difference? What is the difference to Inorder Dominator Tree?
- How do I get the post-dominator tree with your tool? Do you already
have test cases? In fact, normal IR test cases would be really cool. We
do not have a lot of test coverage for all the dominance stuff.
Best,
Tobias
Tobias,
Thanks for the comments and taking a look at my fork.
- It would be convenient to allow .ll files to be read.
Sure, I can add it tomorrow.
- You do not use the BB names in the output, right? This would certainly
improve readability!
You actually don’t have to convert you bitcode files to ‘graph’ files (as long as they contain a single function) – then names should be preserved, but you don’t get to play with updates. The format I’m using isn’t very good, but it’s compatible with some other implementation by Loucas – the author of the algorithm. Do you think that having a format like that in LLVM would make sense? Danny and I though about in the context of quickly writing and modifying tests for dominators and things like the NewGVN.
- My output prints “New Dominator Tree” to times in a row. What is the
difference? What is the difference to Inorder Dominator Tree?
When you graph files have updates (i numFrom numTo and d numFrom numTo) then the first one is the original one and the second one is the one after applying all the updates. Inorder Dominator Tree is the existing DomTree – I used to use it just for visual comparison.
- How do I get the post-dominator tree with your tool? Do you already
have test cases? In fact, normal IR test cases would be really cool. We
do not have a lot of test coverage for all the dominance stuff.
I haven’t really played with postdominators yet. Do you have any specific ideas on how to test it in mind? And I definitely agree, dominators need to be tested more thoroughly. I think that because of the manual updates (of questionable correctness) and frequent recalculations there must be many undiscovered bugs that we just haven’t had a chance to observe yet.
Best,
Kuba
Tobias,
Thanks for the comments and taking a look at my fork.
> - It would be convenient to allow .ll files to be read.
Sure, I can add it tomorrow.
Thank you!
- You do not use the BB names in the output, right? This would certainly
> improve readability!You actually don't have to convert you bitcode files to 'graph' files (as
long as they contain a single function) -- then names should be
preserved,
but you don't get to play with updates. The format I'm using isn't very
good, but it's compatible with some other implementation by Loucas -- the
author of the algorithm.
I tried to use a bitcode file directly:
[grosser@greina0 build]$ cat ../test/Analysis/RegionInfo/test.ll
define void @foo() {
entry:
br i1 1, label %infloop, label %next
infloop:
br label %infloop
next:
ret void
[grosser@greina0 build]$ bin/dominators
../test/Analysis/RegionInfo/test.bc
New Dominator Tree:
[0] %virtual_entry IO{0, 7}, R{nullptr}, P{nullptr}
[1] %entry_n_1 IO{1, 6}, R{virtual_entry}, P{virtual_entry}
[2] %n_2 IO{2, 3}, R{entry_n_1}, P{entry_n_1}
[2] %n_3 IO{4, 5}, R{entry_n_1}, P{entry_n_1}
New Dominator Tree:
[0] %virtual_entry IO{0, 7}, R{nullptr}, P{nullptr}
[1] %entry_n_1 IO{1, 6}, R{virtual_entry}, P{virtual_entry}
[2] %n_2 IO{2, 3}, R{entry_n_1}, P{entry_n_1}
[2] %n_3 IO{4, 5}, R{entry_n_1}, P{entry_n_1}
I think chandler more than me. If me, assume I was wrong
I believe you can prove it is not possible to do this without doing all
the work of the separate passes. The main cost is the DFS walk, and you
can't do an undirected DFS that will work here.
I believe related things have even been proven to show the brokenness of
various algorithms.
See SSI Properties Revisited - Inria section 4.1
Given the difficulties and subtleties here, I would consider any such
approach that tried to share work between the passes as "fraught with peril"
Do you mean the algorithm is slower in debug builds of LLVM, or when
LLVM is used to do debug builds of things?
It sounds like you say the algorithm can be 8% slower for debug
builds, but also likely to speed up debug builds? I'm confused 
Do you mean the algorithm is slower in debug builds of LLVM, or when
LLVM is used to do debug builds of things?It sounds like you say the algorithm can be 8% slower for debug
builds, but also likely to speed up debug builds? I’m confused
The algorithm can be slower when LLVM is built in Debug mode instead of Release.
DomTree is only an analysis, so as long as it is correct it doesn’t affect performance of compiled programs.
> Hi Jakub,
>
> thanks for the update. I was already eagerly waiting to hear what your
> internship on dominance brings. I think the numbers are already very
> encouraging and I believe moving incrementally over to the new algorithm
> is exactly the right approach.
>
> I have a couple of questions:
>
> 1) Daniel and Chandler have for a long time been talking about computing
> dominance and post-dominance in one shot to reduce the cost of
> post-dominance and make it (freely) available everywhere. Is this
> covered by your current (or planned) work?
>I think chandler more than me. If me, assume I was wrong
I believe you can prove it is not possible to do this without doing all
the work of the separate passes. The main cost is the DFS walk, and you
can't do an undirected DFS that will work here.
I believe related things have even been proven to show the brokenness of
various algorithms.
Alright. This sounds to be more in line with what I expected. Thanks for
clarifying.
Best,
Tobias
Hi Kuba,
Hi folks,
This summer I'm working on improving dominators during my internship at
Google. Below is an RFC on switching to dynamic dominators, which you can
also read as a Google Doc if you prefer so. Please let us know what you
think.~Kuba
=======================================================================
1. Context
Currently LLVM uses the Simple Lengauer-Tarjan algorithm to build the
(post)dominator tree. The only way to update it is by manually setting IDoms
which is not obvious in many cases and can be extremely error-prone. And
because updating it manually is so difficult, programmers tend to just
recompute it after changing the CFG (by not AU.addPreserved()'ing the
DomTree). This causes DomTree calculation to fire very often even if only a
very small portion of it gets really affected by the changes in CFG. As an
example, when optimizing a Full LTO clang bitcode, DominatorTreeWrapperPass
alone calls DT.recalculate over 6.5 million times, which takes 20s on my
machine.Using an incremental algorithm it would be much easier to keep an up-to-date
DomTree without custom update logic, which will save us the time currently
spent during DomTree recalculations and reduce the number of bugs caused by
manual updates. It would also make it feasible to maintain postdominators
and use them more broadly, which currently can be too complicated and
expensive.
Another motivating point is that with dominators automatically updated,
the JumpThreading pass can be extended to reason about loops: i.e.,
jump-threading across back-edges of a loop is an important extension
to increase the performance of finite state automata usually implemented
with a loop and a switch stmt within. Brian Rzycki and I have been toying
with keeping the dominators updated across the JumpThreading pass and
that seemed quite involved with the current implementation of domtree.
I'm glad to see work in this area, and we will help testing your prototype
and patches.
Thanks,
Sebastian
Is there a difference here between sharing work and sharing updates? The difficult part of using postdominance seems to be not the one-time cost of computing it, but the fact that we don’t preserve it anywhere. Thus, using it would require re-computing it a lot. Even if we can’t share the work of constructing a dominance and post-dominance tree, could we have an update API that could be used to keep both trees consistent? Thanks again, Hal
+1 -Hal
Hi Jakub,
thanks for the update. I was already eagerly waiting to hear what your
internship on dominance brings. I think the numbers are already very
encouraging and I believe moving incrementally over to the new algorithm
is exactly the right approach.I have a couple of questions:
1) Daniel and Chandler have for a long time been talking about computing
dominance and post-dominance in one shot to reduce the cost of
post-dominance and make it (freely) available everywhere. Is this
covered by your current (or planned) work?I think chandler more than me. If me, assume I was wrong
I believe you can prove it is not possible to do this without doing all
the work of the separate passes. The main cost is the DFS walk, and you
can't do an undirected DFS that will work here.
I believe related things have even been proven to show the brokenness of
various algorithms.See https://hal.inria.fr/hal-00761505 section 4.1
Given the difficulties and subtleties here, I would consider any such
approach that tried to share work between the passes as "fraught with peril"Is there a difference here between sharing work and sharing updates?
Yes.
The difficult part of using postdominance seems to be not the one-time
cost of computing it, but the fact that we don't preserve it anywhere.
Thus, using it would require re-computing it a lot. Even if we can't share
the work of constructing a dominance and post-dominance tree, could we have
an update API that could be used to keep both trees consistent?
This is precisely what Jakub's set of patches solves.
The API for updating either dominance or post-dominance will be identical.
It just needs to be informed of new edges and removed edges in the CFG.
It's possible,at least in the new PM, to easily have a DominatorTreeUpdater
utility that uses getAnalysisIfAvailable on both dominator and
postdominator analysis, and calls through to the update API for each one.
Great! -Hal