[llvm-dev] [PM] I think that the new PM needs to learn about inter-analysis dependencies...

[Sean Silva via llvm-dev]

On Thu, Jul 14, 2016 at 12:51 AM, Sean Silva <chisophugis at gmail.com> wrote:

>
>
> On Wed, Jul 13, 2016 at 1:48 AM, Sean Silva <chisophugis at gmail.com> wrote:
>
>>
>>
>> On Wed, Jul 13, 2016 at 12:34 AM, Chandler Carruth <chandlerc at gmail.com>
>> wrote:
>>
>>> On Wed, Jul 13, 2016 at 12:25 AM Sean Silva <chisophugis at gmail.com>
>>> wrote:
>>>
>>>> On Tue, Jul 12, 2016 at 11:39 PM, Chandler Carruth <chandlerc at gmail.com
>>>> > wrote:
>>>>
>>>>> On Tue, Jul 12, 2016 at 11:34 PM Sean Silva <chisophugis at gmail.com>
>>>>> wrote:
>>>>>
>>>>>> On Tue, Jul 12, 2016 at 11:32 PM, Xinliang David Li <
>>>>>> davidxl at google.com> wrote:
>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On Tue, Jul 12, 2016 at 10:57 PM, Chandler Carruth <
>>>>>>> chandlerc at gmail.com> wrote:
>>>>>>>
>>>>>>>> Yea, this is a nasty problem.
>>>>>>>>
>>>>>>>> One important thing to understand is that this is specific to
>>>>>>>> analyses which hold references to other analyses. While this isn't unheard
>>>>>>>> of, it isn't as common as it could be. Still, definitely something we need
>>>>>>>> to address.
>>>>>>>>
>>>>>>>
>>>>>>> We can call this type of dependencies (holding references)
>>>>>>> hard-dependency. The soft dependency refers to the case where analysis 'A'
>>>>>>> depends on 'B' during computation, but does not need 'B' once it is
>>>>>>> computed.
>>>>>>>
>>>>>>> There are actually quite a few examples of hard-dependency case. For
>>>>>>> instance LoopAccessInfo, LazyValueInfo etc which hold references to other
>>>>>>> analyses.
>>>>>>>
>>>>>>> Problem involving hard-dependency is actually easier to detect, as
>>>>>>> it is usually a compile time problem. Issues involving soft dependencies
>>>>>>> are more subtle and can lead to wrong code gen.
>>>>>>>
>>>>>>
>>>>>> Did you mean to say that soft-dependency problems are easier to
>>>>>> detect? At least my intuition is that soft-dependency is easier because
>>>>>> there is no risk of dangling pointers to other analyses.
>>>>>>
>>>>>
>>>>> The issue is that the fact that there is *any* dependency isn't clear.
>>>>>
>>>>> However, I think the only real problem here are these "hard
>>>>> dependencies" (I don't really like that term though). For others, only an
>>>>> analysis that is *explicitly* preserved survives. So I'm not worried about
>>>>> the fact that people have to remember this.
>>>>>
>>>>> The question is how often there are cross-data-structure references.
>>>>> David mentions a few examples, and I'm sure there are more, but it isn't
>>>>> clear to me yet whether this is pervasive or occasional.
>>>>>
>>>>
>>>>
>>>> I just did a quick run-through of PassRegistry.def and this is what I
>>>> found:
>>>>
>>>> Module analyses: 0/5 hold pointers to other analyses
>>>> CallGraph: No pointers to other analyses.
>>>> LazyCallGraph: No pointers to other analyses.
>>>> ProfileSummaryAnalysis: No pointers to other analyses.
>>>> TargetLibraryAnalysis: No pointers to other analyses.
>>>> VerifierAnalysis: No pointers to other analyses.
>>>>
>>>> Module alias analyses: 1/1 keeps pointer to other analysis.
>>>> GlobalsAA: Result keeps pointer to TLI (this is a function analysis).
>>>>
>>>> Function analyses: 9/17 keep pointers to other analysis
>>>> AAManager: Its Result holds TLI pointer and pointers to individual AA
>>>> result objects.
>>>> AssumptionAnalysis: No pointers to other analyses.
>>>> BlockFrequencyAnalysis: Its Result holds pointers to LoopInfo and BPI.
>>>> BranchProbabilityAnalysis: Stores no pointers to other analyses. (uses
>>>> LoopInfo to "recalculate" though)
>>>> DominatorTreeAnalysis: Stores no pointers to other analyses.
>>>> PostDominatorTreeAnalysis: Stores no pointers to other analyses.
>>>> DemandedBitsAnalysis: Stores pointers to AssumptionCache
>>>> and DominatorTree
>>>> DominanceFrontierAnalysis: Stores no pointers to other analyses.
>>>> (uses DominatorTreeAnalysis for "recalculate" though).
>>>> LoopInfo: Uses DominatorTreeAnalysis for "recalculate" but stores no
>>>> pointers.
>>>> LazyValueAnalysis: Stores pointers to AssumptionCache,
>>>> TargetLibraryInfo, DominatorTree.
>>>> DependenceAnalysis: Stores pointers to AliasAnalysis, ScalarEvolution,
>>>> LoopInfo
>>>> MemoryDependenceAnalysis: Stores pointers to AliasAnalysis,
>>>> AssumptionCache, TargetLibraryInfo, DominatorTree
>>>> MemorySSAAnalysis: Stores pointers to AliasAnalysis, DominatorTree
>>>> RegionInfoAnalysis: Stores pointers to DomTree, PostDomTree, DomFrontier
>>>> ScalarEvolutionAnalysis: Stores pointers to TargetLibraryInfo,
>>>> AssumptionCache, DominatorTree, LoopInfo
>>>> TargetLibraryAnalysis: Has no dependencies
>>>> TargetIRAnalysis: Has no dependencies.
>>>>
>>>> Function alias analyses: 3/5 keep pointers to other analyses
>>>> BasicAA: Keeps pointers to TargetLibraryInfo, AssumptionCache,
>>>> DominatorTree, LoopInfo
>>>> CFLAA: Keeps pointer to TargetLibraryInfo
>>>> SCEVAA: Keeps pointer to ScalarEvolution
>>>> ScopedNoAliasAA: No dependencies
>>>> TypeBasedAA: No dependencies
>>>>
>>>>
>>>> Total: 13/28 analyses (~50%) hold pointers to other analyses.
>>>> Of the 15/28 analyses that don't hold pointers, 12/15 simply have no
>>>> dependencies. Only 3/15 (BPI, LoopInfo, DominanceFrontier) have
>>>> dependencies that are used just for a "recalculate" step that retains no
>>>> pointers.
>>>> So I think it is fair to say that analyses which hold pointers to other
>>>> analyses is not an exceptional case. In fact, analyses that use other
>>>> analyses just for a "recalculate" step seems to be the exceptional case
>>>> (only 3/28 or about 10%)
>>>>
>>>
>>> Interesting!
>>>
>>> Most of these look like they hold a pointer to the root analysis as
>>> opposed to detailed objects *inside* the analysis?
>>>
>>> It might make sense to try to handle this very specific pattern in a
>>> special way of overriding the invalidate routines is too error prone.... We
>>> could try to make this work "automatically" but I'm worried this would be
>>> challenging to get right. Open to suggestions of course.
>>>
>>> Any other ideas about what would make sense to handle this?
>>>
>>> Does it make sense to override the invalidate routines now and iterate
>>> from there? I feel like you've done a lot of the research necessary for
>>> this already...
>>>
>>
>> I'll keep pushing forward tomorrow with building test-suite successfully
>> using the new PM for the LTO pipeline (I was doing some unrelated LLD stuff
>> for most of today). It will be interesting to see how many `invalidate`
>> overrides will be needed to avoid these issues for just the LTO pipeline on
>> test-suite.
>>
>
> I spent the better part of today working on this and will continue
> tomorrow; this problem seems nastier than I thought. For some reason the
> LTO pipeline (or something about LTO) seems to hit on these issues much
> more (I'm talking like 40k lines of ASan error reports from building
> test-suite with the LTO pipeline in the new PM; per-TU steps still using
> the old PM). Some notes:
>
> - BasicAA's dependence on domtree and loopinfo in the new PM seems to
> account for quite a few of the problems.
> - BasicAA and other stuff are marked (by overriding `invalidate` to return
> false) to never be invalidated because they are "stateless". However they
> still hold pointers and so they do need to be invalidated.
> - CallGraph uses AssertingVH (PR28400) and so I needed a workaround
> similar to r274656 in various passes.
> - D21921 is holding up -- I haven't hit any issues with the core logic of
> that patch.
> - AAResults holds handles to various AA result objects. This means it
> pretty much always needs to be invalidated unless you are sure that none of
> the AA's will get invalidated.
>
>
> The existing `invalidate` method doesn't have the right semantics for even
> an error-prone solution :( We are going to need to make some significant
> changes to even get basic sanity I think. Perhaps each analysis can expose
> a "preserve" static function. E.g. instead of `PA.preserve<Foo>();` you
> have to do `Foo::setPreserved(PA);`.
> I'm actually not quite sure that that will even work. Once I have
> test-suite fully building successfully with the LTO pipeline in the new PM
> I'll be able to give a more confident answer (esp. w.r.t. the manager for
> different IRUnitT's).
> But at this point I'm not confident running *any* pass pipeline in the new
> PM without at least assertions+ASan.
>
> We may want to have a proper design discussion around this problem though.
>
> Also I'd like to have test-suite working (by hook or by crook) with LTO in
> the new PM so we can get some numbers on the resident set impact of all
> this caching; if it is really problematic then we may need to start talking
> front-and-center about different invalidation policies for keeping this in
> check instead of leaving it as something that we will be able to patch
> later.
>
>
>
> The more I think about it, the more I'm convinced that the real "hard"
> problem that the new PM is exposing us to is having the ability for any
> pass to ask for any analysis on any IRUnitT (and any specific IRUnit of
> that IRUnitT) and have the result stored somewhere and then invalidated.
> This means that "getAnalysisUsage" is not just a list of passes, but much
> more complicated and is essentially a set of arbitrary pairs "(analysis,
> IRUnit)" (and the associated potential tangle of dependencies between the
> state cached on these tuples). With the old PM, you essentially are looking
> at a problem of scheduling the lifetime of analyses of the same IRUnit
> intermingled with transformation passes on that same IRUnit, so you only
> have the "analysis" part of the tuple above, making things much simpler
> (and handling dependencies is much simpler too).
>

To clarify, it seems like the current new PM is essentially trying to solve
the problem of maintaining/updating a mapping:
(Analysis, IRUnit) -> AnalysisResult
where the AnalysisResult's can have an arbitrary dependency on an arbitrary
set of other AnalysisResult's currently maintained in this mapping. In
order to invalidate any AnalysisResult you need to invalidate all
AnalysisResult's that transitively depend on it. Therefore the right-hand
side of this mapping needs to be something like `(AnalysisResult,
SetOfDependents)`.
So the mapping is really `(Analysis, IRUnit) -> (AnalysisResult,
SetOfDependents)`
Also, this mapping can be updated at any point during the execution of a
transformation pass (and various other places) and must stay correct as the
IR is changed (more on this below).
For example, you might have something like:
(DominatorTreeAnalysis, function @foo) -> (DominatorTree for @foo,
[(DemandedBitsAnalysis, function @foo)])
(AssumptionAnalysis, function @foo) -> (AssumptionCache for @foo,
[(DemandedBitsAnalysis, function @foo)])
(DemandedBitsAnalysis, function @foo) -> (DemandedBits for @foo, [])
(AssumptionAnalysis, function @bar) -> (AssumptionCache for @bar,
[(SomeModuleAnalysis, module TheModule)])
(AssumptionAnalysis, function @baz) -> (AssumptionCache for @baz,
[(SomeModuleAnalysis, module TheModule)])
(SomeModuleAnalysis, module TheModule) -> (SomeModuleAnalysisResult for
TheModule, [(SomeFunctionAnalysis, function @baz)])
(SomeFunctionAnalysis, function @baz) -> (SomeFunctionAnalysisResult for
@baz, [])

So for example, when a transformation pass invalidates
`(AssumptionAnalysis, function @bar)`, we need to walk
`(SomeModuleAnalysis, module TheModule)` and `(SomeFunctionAnalysis,
function @baz)` to invalidate them.


Compare this with the old PM (although like I said we have outgrown this
model). Essentially you take the previous mapping, and require IRUnit to be
a constant at any given point in time. Hence the mapping is essentially
Analysis -> AnalysisResult
Since this is 1:1 there is no real distinction between the Analysis and the
AnalysisResult (and as part of transitioning to the new PM this has had to
be untangled).
This also makes the dependencies simpler since you just have a set of "what
analyses have been run at this point". You just need to run the analyses
individually and make sure they are in the right order. Also, getAnalysis
just takes the Analysis to get the AnalysisResult which makes it simpler --
you just query which analyses are live.


Also, the mapping `(Analysis, IRUnit) -> (AnalysisResult, SetOfDependents)`
that the new PM is essentially trying to keep is even more complicated
because for e.g. Loop and CGSCC passes the IRUnit itself is an object
created by an analysis and subject to invalidation of that analysis as the
IR changes underneath it.

And then there is the question of at what points must this mapping be valid
(i.e. no stale analysis results, no dangling pointers, etc.) and when the
transitive invalidation walking happens. Evidently while a transformation
pass is running, things might temporarily be stale; what are the
"checkpoints" where the mapping is guaranteed to be valid? At the start of
each transformation pass? At least Chandler's D21464 does not stick to this
because the IRUnit's (SCC's) are only updated at the end of running
potentially many function transformation passes. I.e. all but the first
function transformation pass might observe stale IRUnit's (SCC's).

One other thing to note is that soft-dependencies (using David's
terminology) don't require this kind of dependency tracking. An analysis
result can be cached even though its soft-dependencies are not cached. And
invalidation of soft-dependencies does not require invalidating the
soft-dependents. Actually, this makes it the terminology "soft" and "hard'
quite natural; "hard" requires an edge to track the dependency for
invalidation purposes, "soft" does not.

This is all quite general. Perhaps too much. We clearly need to go beyond
the old PM's model, but we may not need to go to the fully general case. Is
there a good middle-ground that meets our needs? What restrictions would we
be willing to live with in order to make it easier? The first one on my
list is to not have the IRUnit's themselves depend on analyses. Like
Chandler mentioned on D21921 this has the effect of e.g. preventing caching
across the intervening module pass in a case like
`module(cgscc(require<foo-cgscc-analysis>),some-module-pass-that-makes-no-changes,cgscc(some-cgscc-pass-that-uses-foo-cgscc-analysis))`
but that seems like a restriction we can live with.


Again, sorry for the braindump.


-- Sean Silva


> We've obviously outgrown this model with examples like LAA,
> AssumptionCacheTracker, etc. that hack around this in the old PM. We may
> want to have a fresh re-examination of what problems we are exactly trying
> to solve.
>
> For me, my main concern now is what changes need to be made in order to
> feel confident running a pipeline in the new PM without assertions+ASan.
>
>
> Sorry for the long post, just brain-dumping before heading home.
>
> -- Sean Silva
>
>
>
>
>>
>> -- Sean Silva
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160714/79ff4bd2/attachment.html>