[llvm-dev] RFC: Getting ProfileSummaryInfo and BlockFrequencyInfo from various types of passes under the new pass manager

[Hiroshi Yamauchi via llvm-dev]

On Mon, Mar 4, 2019 at 10:55 AM Hiroshi Yamauchi <yamauchi at google.com>
wrote:

>
>
> On Sat, Mar 2, 2019 at 12:58 AM Fedor Sergeev <fedor.sergeev at azul.com>
> wrote:
>
>>
>>
>> On 3/2/19 2:38 AM, Hiroshi Yamauchi wrote:
>>
>> Here's a sketch of the proposed approach for just one pass (but imagine
>> more)
>>
>> https://reviews.llvm.org/D58845
>>
>> On Fri, Mar 1, 2019 at 12:54 PM Fedor Sergeev via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>>
>>> On 2/28/19 12:47 AM, Hiroshi Yamauchi via llvm-dev wrote:
>>>
>>> Hi all,
>>>
>>> To implement more profile-guided optimizations, we’d like to use
>>> ProfileSummaryInfo (PSI) and BlockFrequencyInfo (BFI) from more passes of
>>> various types, under the new pass manager.
>>>
>>> The following is what we came up with. Would appreciate feedback. Thanks.
>>>
>>> Issue
>>>
>>> It’s not obvious (to me) how to best do this, given that we cannot
>>> request an outer-scope analysis result from an inner-scope pass through
>>> analysis managers [1] and that we might unnecessarily running some analyses
>>> unless we conditionally build pass pipelines for PGO cases.
>>>
>>> Indeed, this is an intentional restriction in new pass manager, which is
>>> more or less a reflection of a fundamental property of outer-inner IRUnit
>>> relationship
>>> and transformations/analyses run on those units. The main intent for
>>> having those inner IRUnits (e.g. Loops) is to run local transformations and
>>> save compile time
>>> on being local to a particular small piece of IR. Loop Pass manager
>>> allows you to run a whole pipeline of different transformations still
>>> locally, amplifying the save.
>>> As soon as you run function-level analysis from within the loop pipeline
>>> you essentially break this pipelining.
>>> Say, as you run your loop transformation it modifies the loop (and the
>>> function) and potentially invalidates the analysis,
>>> so you have to rerun your analysis again and again. Hence instead of
>>> saving on compile time it ends up increasing it.
>>>
>>
>> Exactly.
>>
>>
>>> I have hit this issue somewhat recently with dependency of loop passes
>>> on BranchProbabilityInfo.
>>> (some loop passes, like IRCE can use it for profitability analysis).
>>>
>> The only solution that appears to be reasonable there is to teach all the
>>> loops passes that need to be pipelined
>>> to preserve BPI (or any other module/function-level analyses) similar to
>>> how they preserve DominatorTree and
>>> other "LoopStandard" analyses.
>>>
>>
>> Is this implemented - do the loop passes preserve BPI?
>>
>> Nope, not implemented right now.
>> One of the problems is that even loop canonicalization passes run at the
>> start of loop pass manager dont preserve it
>> (and at least LoopSimplifyCFG does change control flow).
>>
>>
>> In buildFunctionSimplificationPipeline (where LoopFullUnrollPass is added
>> as in the sketch), LateLoopOptimizationsEPCallbacks
>> and LoopOptimizerEndEPCallbacks seem to allow some arbitrary loop passes to
>> be inserted into the pipelines (via flags)?
>>
>> I wonder how hard it'd be to teach all the relevant loop passes to
>> preserve BFI (or BPI)..
>>
>> Well, each time you restructure control flow around the loops you will
>> have to update those extra analyses,
>> pretty much the same way as DT is being updated through DomTreeUpdater.
>> The trick is to design a proper update interface (and then implement it
>> ;) ).
>> And I have not spent enough time on this issue to get a good idea of what
>> that interface would be.
>>
>
> Hm, sounds non-trivial :) noting BFI depends on BPI.
>

To step back, it looks like:

want to use profiles from more passes -> need to get BFI (from loop passes)
-> need all the loop passes to preserve BFI.

I wonder if there's no way around this.


>
>> regards,
>>   Fedor.
>>
>>
>>> It seems that for different types of passes to be able to get PSI and
>>> BFI, we’d need to ensure PSI is cached for a non-module pass, and PSI, BFI
>>> and the ModuleAnalysisManager proxy are cached for a loop pass in the pass
>>> pipelines. This may mean potentially needing to insert BFI/PSI in front of
>>> many passes [2]. It seems not obvious how to conditionally insert BFI for
>>> PGO pipelines because there isn’t always a good flag to detect PGO cases
>>> [3] or we tend to build pass pipelines before examining the code (or
>>> without propagating enough info down) [4].
>>>
>>> Proposed approach
>>>
>>> - Cache PSI right after the profile summary in the IR is written in the
>>> pass pipeline [5]. This would avoid the need to insert RequiredAnalysisPass
>>> for PSI before each non-module pass that needs it. PSI can be technically
>>> invalidated but unlikely. If it does, we insert another RequiredAnalysisPass
>>>  [6].
>>>
>>> - Conditionally insert RequireAnalysisPass for BFI, if PGO, right before
>>> each loop pass that needs it. This doesn't seem avoidable because BFI can
>>> be invalidated whenever the CFG changes. We detect PGO based on the command
>>> line flags and/or whether the module has the profile summary info (we
>>> may need to pass the module to more functions.)
>>>
>>> - Add a new proxy ModuleAnalysisManagerLoopProxy for a loop pass to be
>>> able to get to the ModuleAnalysisManager in one step and PSI through it.
>>>
>>> Alternative approaches
>>>
>>> Dropping BFI and use PSI only
>>> We could consider not using BFI and solely relying on PSI and
>>> function-level profiles only (as opposed to block-level), but profile
>>> precision would suffer.
>>>
>>> Computing BFI in-place
>>> We could consider computing BFI “in-place” by directly running BFI
>>> outside of the pass manager [7]. This would let us avoid using the analysis
>>> manager constraints but it would still involve running an outer-scope
>>> analysis from an inner-scope pass and potentially cause problems in terms
>>> of pass pipelining and concurrency. Moreover, a potential downside of
>>> running analyses in-place is that it won’t take advantage of cached
>>> analysis results provided by the pass manager.
>>>
>>> Adding inner-scope versions of PSI and BFI
>>> We could consider adding a function-level and loop-level PSI and
>>> loop-level BFI, which internally act like their outer-scope versions but
>>> provide inner-scope results only. This way, we could always call getResult
>>> for PSI and BFI. However, this would still involve running an outer-scope
>>> analysis from an inner-scope pass.
>>>
>>> Caching the FAM and the MAM proxies
>>> We could consider caching the FunctionalAnalysisManager and the
>>> ModuleAnalysisManager proxies once early on instead of adding a new proxy.
>>> But it seems to not likely work well because the analysis cache key type
>>> includes the function or the module and some pass may add a new function
>>> for which the proxy wouldn’t be cached. We’d need to write and insert a
>>> pass in select locations to just fill the cache. Adding the new proxy would
>>> take care of these with a three-line change.
>>>
>>> Conditional BFI
>>> We could consider adding a conditional BFI analysis that is a wrapper
>>> around BFI and computes BFI only if profiles are available (either checking
>>> the module has profile summary or depend on the PSI.) With this, we
>>> wouldn’t need to conditionally build pass pipelines and may work for the
>>> new pass manager. But a similar wouldn’t work for the old pass manager
>>> because we cannot conditionally depend on an analysis under it.
>>>
>>> There is LazyBlockFrequencyInfo.
>>> Not sure how well it fits this idea.
>>>
>>
>> Good point. LazyBlockFrequencyInfo seems usable with the old pass manager
>> (save unnecessary BFI/BPI) and would work for function passes. I think t
>> he restriction still applies - a loop pass cannot still request
>> (outer-scope) BFI, lazy or not, new or old (pass manager). Another
>> assumption is that it'd be cheap and safe to unconditionally depend on
>> PSI or check the module's profile summary.
>>
>>
>>> regards,
>>>   Fedor.
>>>
>>>
>>>
>>> [1] We cannot call AnalysisManager::getResult for an outer scope but
>>> only getCachedResult. Probably because of potential pipelining or
>>> concurrency issues.
>>> [2] For example, potentially breaking up multiple pipelined loop passes
>>> and insert RequireAnalysisPass<BlockFrequencyAnalysis> in front of each of
>>> them.
>>> [3] For example, -fprofile-instr-use and -fprofile-sample-use aren’t
>>> present in ThinLTO post link builds.
>>> [4] For example, we could check whether the module has the profile
>>> summary metadata annotated when building pass pipelines but we don’t always
>>> pass the module down to the place where we build pass pipelines.
>>> [5] By inserting RequireAnalysisPass<ProfileSummaryInfo> after the
>>> PGOInstrumentationUse and the SampleProfileLoaderPass passes (and around
>>> the PGOIndirectCallPromotion pass for the Thin LTO post link pipeline.)
>>> [6] For example, the context-sensitive PGO.
>>> [7] Directly calling its constructor along with the dependent analyses
>>> results, eg. the jump threading pass.
>>>
>>> _______________________________________________
>>> LLVM Developers mailing listllvm-dev at lists.llvm.orghttps://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>
>>>
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org
>>> https://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/20190304/b0d19089/attachment.html>