Hiroshi Yamauchi via llvm-dev
2019-Mar-04 18:55 UTC
[llvm-dev] RFC: Getting ProfileSummaryInfo and BlockFrequencyInfo from various types of passes under the new pass manager
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.> 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/20190304/92803f78/attachment.html>
Hiroshi Yamauchi via llvm-dev
2019-Mar-04 19:49 UTC
[llvm-dev] RFC: Getting ProfileSummaryInfo and BlockFrequencyInfo from various types of passes under the new pass manager
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>
Fedor Sergeev via llvm-dev
2019-Mar-04 22:05 UTC
[llvm-dev] RFC: Getting ProfileSummaryInfo and BlockFrequencyInfo from various types of passes under the new pass manager
On 3/4/19 10:49 PM, Hiroshi Yamauchi wrote:> > > On Mon, Mar 4, 2019 at 10:55 AM Hiroshi Yamauchi <yamauchi at google.com > <mailto:yamauchi at google.com>> wrote: > > > > On Sat, Mar 2, 2019 at 12:58 AM Fedor Sergeev > <fedor.sergeev at azul.com <mailto: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 <mailto: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.Indeed. I believe this is a general consensus here. regards, Fedor.> > > > 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 the 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 list >>> llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> >>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >> >> _______________________________________________ >> LLVM Developers mailing list >> llvm-dev at lists.llvm.org <mailto: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/20190305/054eda98/attachment.html>
Seemingly Similar Threads
- RFC: Getting ProfileSummaryInfo and BlockFrequencyInfo from various types of passes under the new pass manager
- RFC: Getting ProfileSummaryInfo and BlockFrequencyInfo from various types of passes under the new pass manager
- RFC: Getting ProfileSummaryInfo and BlockFrequencyInfo from various types of passes under the new pass manager
- RFC: Getting ProfileSummaryInfo and BlockFrequencyInfo from various types of passes under the new pass manager
- RFC: Getting ProfileSummaryInfo and BlockFrequencyInfo from various types of passes under the new pass manager