llvm-project/llvm/lib/Analysis/CGSCCPassManager.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

1250 lines
51 KiB
C++
Raw Normal View History

//===- CGSCCPassManager.cpp - Managing & running CGSCC passes -------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
#include "llvm/Analysis/CGSCCPassManager.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/Optional.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/iterator_range.h"
#include "llvm/Analysis/LazyCallGraph.h"
#include "llvm/IR/Constant.h"
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/PassManager.h"
Add PassManagerImpl.h to hide implementation details ClangBuildAnalyzer results show that a lot of time is spent instantiating AnalysisManager::getResultImpl across the code base: **** Templates that took longest to instantiate: 50445 ms: llvm::AnalysisManager<llvm::Function>::getResultImpl (412 times, avg 122 ms) 47797 ms: llvm::AnalysisManager<llvm::Function>::getResult<llvm::TargetLibraryAnalysis> (389 times, avg 122 ms) 46894 ms: std::tie<const unsigned long long, const bool> (2452 times, avg 19 ms) 43851 ms: llvm::BumpPtrAllocatorImpl<llvm::MallocAllocator, 4096, 4096>::Allocate (3228 times, avg 13 ms) 33911 ms: std::tie<const unsigned int, const unsigned int, const unsigned int, const unsigned int> (897 times, avg 37 ms) 33854 ms: std::tie<const unsigned long long, const unsigned long long> (1897 times, avg 17 ms) 27886 ms: std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string (11156 times, avg 2 ms) I mentioned this result to @chandlerc, and he suggested this direction. AnalysisManager is already explicitly instantiated, and getResultImpl doesn't need to be inlined. Move the definition to an Impl header, and include that header in files that explicitly instantiate AnalysisManager. There are only four (real) IR units: - function - module - loop - cgscc Looking at a specific transform (ArgumentPromotion.cpp), here are three compilations before & after this change: BEFORE: $ for i in $(seq 3) ; do ./ccit.bat ; done peak memory: 258.15MB real: 0m6.297s peak memory: 257.54MB real: 0m5.906s peak memory: 257.47MB real: 0m6.219s AFTER: $ for i in $(seq 3) ; do ./ccit.bat ; done peak memory: 235.35MB real: 0m5.454s peak memory: 234.72MB real: 0m5.235s peak memory: 234.39MB real: 0m5.469s The 20MB of memory saved seems real, and the time improvement seems like it is there. Reviewed By: MaskRay Differential Revision: https://reviews.llvm.org/D73817
2020-02-01 08:05:32 +08:00
#include "llvm/IR/PassManagerImpl.h"
Reland [CGSCC] Detect devirtualization in more cases The devirtualization wrapper misses cases where if it wraps a pass manager, an individual pass may devirtualize an indirect call created by a previous pass. For example, inlining may create a new indirect call which is devirtualized by instcombine. Currently the devirtualization wrapper will not see that because it only checks cgscc edges at the very beginning and end of the pass (manager) it wraps. This fixes some tests testing this exact behavior in the legacy PM. Instead of checking WeakTrackingVHs for CallBases at the very beginning and end of the pass it wraps, check every time updateCGAndAnalysisManagerForPass() is called. check-llvm and check-clang with -abort-on-max-devirt-iterations-reached on by default doesn't show any failures outside of tests specifically testing it so it doesn't needlessly rerun passes more than necessary. (The NPM -O2/3 pipeline run the inliner/function simplification pipeline under a devirtualization repeater pass up to 4 times by default). http://llvm-compile-time-tracker.com/?config=O3&stat=instructions&remote=aeubanks shows that 7zip has ~1% compile time regression. I looked at it and saw that there indeed was devirtualization happening that was not previously caught, so now it reruns the CGSCC pipeline on some SCCs, which is WAI. The initial land assumed CallBase WeakTrackingVHs would always be CallBases, but they can be RAUW'd with undef. Reviewed By: asbirlea Differential Revision: https://reviews.llvm.org/D89587
2020-10-15 08:56:38 +08:00
#include "llvm/IR/ValueHandle.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/TimeProfiler.h"
#include "llvm/Support/raw_ostream.h"
#include <algorithm>
#include <cassert>
#include <iterator>
#define DEBUG_TYPE "cgscc"
using namespace llvm;
// Explicit template instantiations and specialization definitions for core
[PM] Support invalidation of inner analysis managers from a pass over the outer IR unit. Summary: This never really got implemented, and was very hard to test before a lot of the refactoring changes to make things more robust. But now we can test it thoroughly and cleanly, especially at the CGSCC level. The core idea is that when an inner analysis manager proxy receives the invalidation event for the outer IR unit, it needs to walk the inner IR units and propagate it to the inner analysis manager for each of those units. For example, each function in the SCC needs to get an invalidation event when the SCC gets one. The function / module interaction is somewhat boring here. This really becomes interesting in the face of analysis-backed IR units. This patch effectively handles all of the CGSCC layer's needs -- both invalidating SCC analysis and invalidating function analysis when an SCC gets invalidated. However, this second aspect doesn't really handle the LoopAnalysisManager well at this point. That one will need some change of design in order to fully integrate, because unlike the call graph, the entire function behind a LoopAnalysis's results can vanish out from under us, and we won't even have a cached API to access. I'd like to try to separate solving the loop problems into a subsequent patch though in order to keep this more focused so I've adapted them to the API and updated the tests that immediately fail, but I've not added the level of testing and validation at that layer that I have at the CGSCC layer. An important aspect of this change is that the proxy for the FunctionAnalysisManager at the SCC pass layer doesn't work like the other proxies for an inner IR unit as it doesn't directly manage the FunctionAnalysisManager and invalidation or clearing of it. This would create an ever worsening problem of dual ownership of this responsibility, split between the module-level FAM proxy and this SCC-level FAM proxy. Instead, this patch changes the SCC-level FAM proxy to work in terms of the module-level proxy and defer to it to handle much of the updates. It only does SCC-specific invalidation. This will become more important in subsequent patches that support more complex invalidaiton scenarios. Reviewers: jlebar Subscribers: mehdi_amini, mcrosier, mzolotukhin, llvm-commits Differential Revision: https://reviews.llvm.org/D27197 llvm-svn: 289317
2016-12-10 14:34:44 +08:00
// template typedefs.
namespace llvm {
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
static cl::opt<bool> AbortOnMaxDevirtIterationsReached(
"abort-on-max-devirt-iterations-reached",
cl::desc("Abort when the max iterations for devirtualization CGSCC repeat "
"pass is reached"));
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// Explicit instantiations for the core proxy templates.
[PM] Extend the explicit 'invalidate' method API on analysis results to accept an Invalidator that allows them to invalidate themselves if their dependencies are in turn invalidated. Rather than recording the dependency graph ahead of time when analysis get results from other analyses, this simply lets each result trigger the immediate invalidation of any analyses they actually depend on. They do this in a way that has three nice properties: 1) They don't have to handle transitive dependencies because the infrastructure will recurse for them. 2) The invalidate methods are still called only once. We just dynamically discover the necessary topological ordering, everything is memoized nicely. 3) The infrastructure still provides a default implementation and can access it so that only analyses which have dependencies need to do anything custom. To make this work at all, the invalidation logic also has to defer the deletion of the result objects themselves so that they can remain alive until we have collected the complete set of results to invalidate. A unittest is added here that has exactly the dependency pattern we are concerned with. It hit the use-after-free described by Sean in much detail in the long thread about analysis invalidation before this change, and even in an intermediate form of this change where we failed to defer the deletion of the result objects. There is an important problem with doing dependency invalidation that *isn't* solved here: we don't *enforce* that results correctly invalidate all the analyses whose results they depend on. I actually looked at what it would take to do that, and it isn't as hard as I had thought but the complexity it introduces seems very likely to outweigh the benefit. The technique would be to provide a base class for an analysis result that would be populated with other results, and automatically provide the invalidate method which immediately does the correct thing. This approach has some nice pros IMO: - Handles the case we care about and nothing else: only *results* that depend on other analyses trigger extra invalidation. - Localized to the result rather than centralized in the analysis manager. - Ties the storage of the reference to another result to the triggering of the invalidation of that analysis. - Still supports extending invalidation in customized ways. But the down sides here are: - Very heavy-weight meta-programming is needed to provide this base class. - Requires a pretty awful API for accessing the dependencies. Ultimately, I fear it will not pull its weight. But we can re-evaluate this at any point if we start discovering consistent problems where the invalidation and dependencies get out of sync. It will fit as a clean layer on top of the facilities in this patch that we can add if and when we need it. Note that I'm not really thrilled with the names for these APIs... The name "Invalidator" seems ok but not great. The method name "invalidate" also. In review some improvements were suggested, but they really need *other* uses of these terms to be updated as well so I'm going to do that in a follow-up commit. I'm working on the actual fixes to various analyses that need to use these, but I want to try to get tests for each of them so we don't regress. And those changes are seperable and obvious so once this goes in I should be able to roll them out throughout LLVM. Many thanks to Sean, Justin, and others for help reviewing here. Differential Revision: https://reviews.llvm.org/D23738 llvm-svn: 288077
2016-11-29 06:04:31 +08:00
template class AllAnalysesOn<LazyCallGraph::SCC>;
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
template class AnalysisManager<LazyCallGraph::SCC, LazyCallGraph &>;
template class PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager,
LazyCallGraph &, CGSCCUpdateResult &>;
template class InnerAnalysisManagerProxy<CGSCCAnalysisManager, Module>;
template class OuterAnalysisManagerProxy<ModuleAnalysisManager,
LazyCallGraph::SCC, LazyCallGraph &>;
template class OuterAnalysisManagerProxy<CGSCCAnalysisManager, Function>;
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
/// Explicitly specialize the pass manager run method to handle call graph
/// updates.
template <>
PreservedAnalyses
PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager, LazyCallGraph &,
CGSCCUpdateResult &>::run(LazyCallGraph::SCC &InitialC,
CGSCCAnalysisManager &AM,
LazyCallGraph &G, CGSCCUpdateResult &UR) {
[New PM] Introducing PassInstrumentation framework Pass Execution Instrumentation interface enables customizable instrumentation of pass execution, as per "RFC: Pass Execution Instrumentation interface" posted 06/07/2018 on llvm-dev@ The intent is to provide a common machinery to implement all the pass-execution-debugging features like print-before/after, opt-bisect, time-passes etc. Here we get a basic implementation consisting of: * PassInstrumentationCallbacks class that handles registration of callbacks and access to them. * PassInstrumentation class that handles instrumentation-point interfaces that call into PassInstrumentationCallbacks. * Callbacks accept StringRef which is just a name of the Pass right now. There were some ideas to pass an opaque wrapper for the pointer to pass instance, however it appears that pointer does not actually identify the instance (adaptors and managers might have the same address with the pass they govern). Hence it was decided to go simple for now and then later decide on what the proper mental model of identifying a "pass in a phase of pipeline" is. * Callbacks accept llvm::Any serving as a wrapper for const IRUnit*, to remove direct dependencies on different IRUnits (e.g. Analyses). * PassInstrumentationAnalysis analysis is explicitly requested from PassManager through usual AnalysisManager::getResult. All pass managers were updated to run that to get PassInstrumentation object for instrumentation calls. * Using tuples/index_sequence getAnalysisResult helper to extract generic AnalysisManager's extra args out of a generic PassManager's extra args. This is the only way I was able to explicitly run getResult for PassInstrumentationAnalysis out of a generic code like PassManager::run or RepeatedPass::run. TODO: Upon lengthy discussions we agreed to accept this as an initial implementation and then get rid of getAnalysisResult by improving RepeatedPass implementation. * PassBuilder takes PassInstrumentationCallbacks object to pass it further into PassInstrumentationAnalysis. Callbacks registration should be performed directly through PassInstrumentationCallbacks. * new-pm tests updated to account for PassInstrumentationAnalysis being run * Added PassInstrumentation tests to PassBuilderCallbacks unit tests. Other unit tests updated with registration of the now-required PassInstrumentationAnalysis. Made getName helper to return std::string (instead of StringRef initially) to fix asan builtbot failures on CGSCC tests. Reviewers: chandlerc, philip.pfaffe Differential Revision: https://reviews.llvm.org/D47858 llvm-svn: 342664
2018-09-21 01:08:45 +08:00
// Request PassInstrumentation from analysis manager, will use it to run
// instrumenting callbacks for the passes later.
PassInstrumentation PI =
AM.getResult<PassInstrumentationAnalysis>(InitialC, G);
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
PreservedAnalyses PA = PreservedAnalyses::all();
if (DebugLogging)
dbgs() << "Starting CGSCC pass manager run.\n";
// The SCC may be refined while we are running passes over it, so set up
// a pointer that we can update.
LazyCallGraph::SCC *C = &InitialC;
// Get Function analysis manager from its proxy.
FunctionAnalysisManager &FAM =
AM.getCachedResult<FunctionAnalysisManagerCGSCCProxy>(*C)->getManager();
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
for (auto &Pass : Passes) {
[New PM] Introducing PassInstrumentation framework Pass Execution Instrumentation interface enables customizable instrumentation of pass execution, as per "RFC: Pass Execution Instrumentation interface" posted 06/07/2018 on llvm-dev@ The intent is to provide a common machinery to implement all the pass-execution-debugging features like print-before/after, opt-bisect, time-passes etc. Here we get a basic implementation consisting of: * PassInstrumentationCallbacks class that handles registration of callbacks and access to them. * PassInstrumentation class that handles instrumentation-point interfaces that call into PassInstrumentationCallbacks. * Callbacks accept StringRef which is just a name of the Pass right now. There were some ideas to pass an opaque wrapper for the pointer to pass instance, however it appears that pointer does not actually identify the instance (adaptors and managers might have the same address with the pass they govern). Hence it was decided to go simple for now and then later decide on what the proper mental model of identifying a "pass in a phase of pipeline" is. * Callbacks accept llvm::Any serving as a wrapper for const IRUnit*, to remove direct dependencies on different IRUnits (e.g. Analyses). * PassInstrumentationAnalysis analysis is explicitly requested from PassManager through usual AnalysisManager::getResult. All pass managers were updated to run that to get PassInstrumentation object for instrumentation calls. * Using tuples/index_sequence getAnalysisResult helper to extract generic AnalysisManager's extra args out of a generic PassManager's extra args. This is the only way I was able to explicitly run getResult for PassInstrumentationAnalysis out of a generic code like PassManager::run or RepeatedPass::run. TODO: Upon lengthy discussions we agreed to accept this as an initial implementation and then get rid of getAnalysisResult by improving RepeatedPass implementation. * PassBuilder takes PassInstrumentationCallbacks object to pass it further into PassInstrumentationAnalysis. Callbacks registration should be performed directly through PassInstrumentationCallbacks. * new-pm tests updated to account for PassInstrumentationAnalysis being run * Added PassInstrumentation tests to PassBuilderCallbacks unit tests. Other unit tests updated with registration of the now-required PassInstrumentationAnalysis. Made getName helper to return std::string (instead of StringRef initially) to fix asan builtbot failures on CGSCC tests. Reviewers: chandlerc, philip.pfaffe Differential Revision: https://reviews.llvm.org/D47858 llvm-svn: 342664
2018-09-21 01:08:45 +08:00
// Check the PassInstrumentation's BeforePass callbacks before running the
// pass, skip its execution completely if asked to (callback returns false).
if (!PI.runBeforePass(*Pass, *C))
continue;
PreservedAnalyses PassPA;
{
TimeTraceScope TimeScope(Pass->name());
PassPA = Pass->run(*C, AM, G, UR);
}
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
if (UR.InvalidatedSCCs.count(C))
PI.runAfterPassInvalidated<LazyCallGraph::SCC>(*Pass, PassPA);
else
PI.runAfterPass<LazyCallGraph::SCC>(*Pass, *C, PassPA);
[New PM] Introducing PassInstrumentation framework Pass Execution Instrumentation interface enables customizable instrumentation of pass execution, as per "RFC: Pass Execution Instrumentation interface" posted 06/07/2018 on llvm-dev@ The intent is to provide a common machinery to implement all the pass-execution-debugging features like print-before/after, opt-bisect, time-passes etc. Here we get a basic implementation consisting of: * PassInstrumentationCallbacks class that handles registration of callbacks and access to them. * PassInstrumentation class that handles instrumentation-point interfaces that call into PassInstrumentationCallbacks. * Callbacks accept StringRef which is just a name of the Pass right now. There were some ideas to pass an opaque wrapper for the pointer to pass instance, however it appears that pointer does not actually identify the instance (adaptors and managers might have the same address with the pass they govern). Hence it was decided to go simple for now and then later decide on what the proper mental model of identifying a "pass in a phase of pipeline" is. * Callbacks accept llvm::Any serving as a wrapper for const IRUnit*, to remove direct dependencies on different IRUnits (e.g. Analyses). * PassInstrumentationAnalysis analysis is explicitly requested from PassManager through usual AnalysisManager::getResult. All pass managers were updated to run that to get PassInstrumentation object for instrumentation calls. * Using tuples/index_sequence getAnalysisResult helper to extract generic AnalysisManager's extra args out of a generic PassManager's extra args. This is the only way I was able to explicitly run getResult for PassInstrumentationAnalysis out of a generic code like PassManager::run or RepeatedPass::run. TODO: Upon lengthy discussions we agreed to accept this as an initial implementation and then get rid of getAnalysisResult by improving RepeatedPass implementation. * PassBuilder takes PassInstrumentationCallbacks object to pass it further into PassInstrumentationAnalysis. Callbacks registration should be performed directly through PassInstrumentationCallbacks. * new-pm tests updated to account for PassInstrumentationAnalysis being run * Added PassInstrumentation tests to PassBuilderCallbacks unit tests. Other unit tests updated with registration of the now-required PassInstrumentationAnalysis. Made getName helper to return std::string (instead of StringRef initially) to fix asan builtbot failures on CGSCC tests. Reviewers: chandlerc, philip.pfaffe Differential Revision: https://reviews.llvm.org/D47858 llvm-svn: 342664
2018-09-21 01:08:45 +08:00
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// Update the SCC if necessary.
C = UR.UpdatedC ? UR.UpdatedC : C;
if (UR.UpdatedC) {
// If C is updated, also create a proxy and update FAM inside the result.
auto *ResultFAMCP =
&AM.getResult<FunctionAnalysisManagerCGSCCProxy>(*C, G);
ResultFAMCP->updateFAM(FAM);
}
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// If the CGSCC pass wasn't able to provide a valid updated SCC, the
// current SCC may simply need to be skipped if invalid.
if (UR.InvalidatedSCCs.count(C)) {
LLVM_DEBUG(dbgs() << "Skipping invalidated root or island SCC!\n");
break;
}
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// Check that we didn't miss any update scenario.
assert(C->begin() != C->end() && "Cannot have an empty SCC!");
// Update the analysis manager as each pass runs and potentially
// invalidates analyses.
AM.invalidate(*C, PassPA);
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// Finally, we intersect the final preserved analyses to compute the
// aggregate preserved set for this pass manager.
PA.intersect(std::move(PassPA));
// FIXME: Historically, the pass managers all called the LLVM context's
// yield function here. We don't have a generic way to acquire the
// context and it isn't yet clear what the right pattern is for yielding
// in the new pass manager so it is currently omitted.
// ...getContext().yield();
}
// Before we mark all of *this* SCC's analyses as preserved below, intersect
// this with the cross-SCC preserved analysis set. This is used to allow
// CGSCC passes to mutate ancestor SCCs and still trigger proper invalidation
// for them.
UR.CrossSCCPA.intersect(PA);
// Invalidation was handled after each pass in the above loop for the current
// SCC. Therefore, the remaining analysis results in the AnalysisManager are
// preserved. We mark this with a set so that we don't need to inspect each
// one individually.
[PM] Introduce the facilities for registering cross-IR-unit dependencies that require deferred invalidation. This handles the other real-world invalidation scenario that we have cases of: a function analysis which caches references to a module analysis. We currently do this in the AA aggregation layer and might well do this in other places as well. Since this is relative rare, the technique is somewhat more cumbersome. Analyses need to register themselves when accessing the outer analysis manager's proxy. This proxy is already necessarily present to allow access to the outer IR unit's analyses. By registering here we can track and trigger invalidation when that outer analysis goes away. To make this work we need to enhance the PreservedAnalyses infrastructure to support a (slightly) more explicit model for "sets" of analyses, and allow abandoning a single specific analyses even when a set covering that analysis is preserved. That allows us to describe the scenario of preserving all Function analyses *except* for the one where deferred invalidation has triggered. We also need to teach the invalidator API to support direct ID calls instead of always going through a template to dispatch so that we can just record the ID mapping. I've introduced testing of all of this both for simple module<->function cases as well as for more complex cases involving a CGSCC layer. Much like the previous patch I've not tried to fully update the loop pass management layer because that layer is due to be heavily reworked to use similar techniques to the CGSCC to handle updates. As that happens, we'll have a better testing basis for adding support like this. Many thanks to both Justin and Sean for the extensive reviews on this to help bring the API design and documentation into a better state. Differential Revision: https://reviews.llvm.org/D27198 llvm-svn: 290594
2016-12-27 16:40:39 +08:00
PA.preserveSet<AllAnalysesOn<LazyCallGraph::SCC>>();
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
if (DebugLogging)
dbgs() << "Finished CGSCC pass manager run.\n";
return PA;
}
PreservedAnalyses
ModuleToPostOrderCGSCCPassAdaptor::run(Module &M, ModuleAnalysisManager &AM) {
// Setup the CGSCC analysis manager from its proxy.
CGSCCAnalysisManager &CGAM =
AM.getResult<CGSCCAnalysisManagerModuleProxy>(M).getManager();
// Get the call graph for this module.
LazyCallGraph &CG = AM.getResult<LazyCallGraphAnalysis>(M);
// Get Function analysis manager from its proxy.
FunctionAnalysisManager &FAM =
AM.getCachedResult<FunctionAnalysisManagerModuleProxy>(M)->getManager();
// We keep worklists to allow us to push more work onto the pass manager as
// the passes are run.
SmallPriorityWorklist<LazyCallGraph::RefSCC *, 1> RCWorklist;
SmallPriorityWorklist<LazyCallGraph::SCC *, 1> CWorklist;
// Keep sets for invalidated SCCs and RefSCCs that should be skipped when
// iterating off the worklists.
SmallPtrSet<LazyCallGraph::RefSCC *, 4> InvalidRefSCCSet;
SmallPtrSet<LazyCallGraph::SCC *, 4> InvalidSCCSet;
SmallDenseSet<std::pair<LazyCallGraph::Node *, LazyCallGraph::SCC *>, 4>
InlinedInternalEdges;
CGSCCUpdateResult UR = {
RCWorklist, CWorklist, InvalidRefSCCSet, InvalidSCCSet,
nullptr, nullptr, PreservedAnalyses::all(), InlinedInternalEdges,
{}};
// Request PassInstrumentation from analysis manager, will use it to run
// instrumenting callbacks for the passes later.
PassInstrumentation PI = AM.getResult<PassInstrumentationAnalysis>(M);
PreservedAnalyses PA = PreservedAnalyses::all();
CG.buildRefSCCs();
for (auto RCI = CG.postorder_ref_scc_begin(),
RCE = CG.postorder_ref_scc_end();
RCI != RCE;) {
assert(RCWorklist.empty() &&
"Should always start with an empty RefSCC worklist");
// The postorder_ref_sccs range we are walking is lazily constructed, so
// we only push the first one onto the worklist. The worklist allows us
// to capture *new* RefSCCs created during transformations.
//
// We really want to form RefSCCs lazily because that makes them cheaper
// to update as the program is simplified and allows us to have greater
// cache locality as forming a RefSCC touches all the parts of all the
// functions within that RefSCC.
//
// We also eagerly increment the iterator to the next position because
// the CGSCC passes below may delete the current RefSCC.
RCWorklist.insert(&*RCI++);
do {
LazyCallGraph::RefSCC *RC = RCWorklist.pop_back_val();
if (InvalidRefSCCSet.count(RC)) {
LLVM_DEBUG(dbgs() << "Skipping an invalid RefSCC...\n");
continue;
}
assert(CWorklist.empty() &&
"Should always start with an empty SCC worklist");
LLVM_DEBUG(dbgs() << "Running an SCC pass across the RefSCC: " << *RC
<< "\n");
// The top of the worklist may *also* be the same SCC we just ran over
// (and invalidated for). Keep track of that last SCC we processed due
// to SCC update to avoid redundant processing when an SCC is both just
// updated itself and at the top of the worklist.
LazyCallGraph::SCC *LastUpdatedC = nullptr;
// Push the initial SCCs in reverse post-order as we'll pop off the
// back and so see this in post-order.
for (LazyCallGraph::SCC &C : llvm::reverse(*RC))
CWorklist.insert(&C);
do {
LazyCallGraph::SCC *C = CWorklist.pop_back_val();
// Due to call graph mutations, we may have invalid SCCs or SCCs from
// other RefSCCs in the worklist. The invalid ones are dead and the
// other RefSCCs should be queued above, so we just need to skip both
// scenarios here.
if (InvalidSCCSet.count(C)) {
LLVM_DEBUG(dbgs() << "Skipping an invalid SCC...\n");
continue;
}
if (LastUpdatedC == C) {
LLVM_DEBUG(dbgs() << "Skipping redundant run on SCC: " << *C << "\n");
continue;
}
if (&C->getOuterRefSCC() != RC) {
LLVM_DEBUG(dbgs() << "Skipping an SCC that is now part of some other "
"RefSCC...\n");
continue;
}
// Ensure we can proxy analysis updates from the CGSCC analysis manager
// into the the Function analysis manager by getting a proxy here.
// This also needs to update the FunctionAnalysisManager, as this may be
// the first time we see this SCC.
CGAM.getResult<FunctionAnalysisManagerCGSCCProxy>(*C, CG).updateFAM(
FAM);
// Each time we visit a new SCC pulled off the worklist,
// a transformation of a child SCC may have also modified this parent
// and invalidated analyses. So we invalidate using the update record's
// cross-SCC preserved set. This preserved set is intersected by any
// CGSCC pass that handles invalidation (primarily pass managers) prior
// to marking its SCC as preserved. That lets us track everything that
// might need invalidation across SCCs without excessive invalidations
// on a single SCC.
//
// This essentially allows SCC passes to freely invalidate analyses
// of any ancestor SCC. If this becomes detrimental to successfully
// caching analyses, we could force each SCC pass to manually
// invalidate the analyses for any SCCs other than themselves which
// are mutated. However, that seems to lose the robustness of the
// pass-manager driven invalidation scheme.
CGAM.invalidate(*C, UR.CrossSCCPA);
do {
// Check that we didn't miss any update scenario.
assert(!InvalidSCCSet.count(C) && "Processing an invalid SCC!");
assert(C->begin() != C->end() && "Cannot have an empty SCC!");
assert(&C->getOuterRefSCC() == RC &&
"Processing an SCC in a different RefSCC!");
LastUpdatedC = UR.UpdatedC;
UR.UpdatedRC = nullptr;
UR.UpdatedC = nullptr;
// Check the PassInstrumentation's BeforePass callbacks before
// running the pass, skip its execution completely if asked to
// (callback returns false).
if (!PI.runBeforePass<LazyCallGraph::SCC>(*Pass, *C))
continue;
PreservedAnalyses PassPA;
{
TimeTraceScope TimeScope(Pass->name());
PassPA = Pass->run(*C, CGAM, CG, UR);
}
if (UR.InvalidatedSCCs.count(C))
PI.runAfterPassInvalidated<LazyCallGraph::SCC>(*Pass, PassPA);
else
PI.runAfterPass<LazyCallGraph::SCC>(*Pass, *C, PassPA);
// Update the SCC and RefSCC if necessary.
C = UR.UpdatedC ? UR.UpdatedC : C;
RC = UR.UpdatedRC ? UR.UpdatedRC : RC;
if (UR.UpdatedC) {
// If we're updating the SCC, also update the FAM inside the proxy's
// result.
CGAM.getResult<FunctionAnalysisManagerCGSCCProxy>(*C, CG).updateFAM(
FAM);
}
// If the CGSCC pass wasn't able to provide a valid updated SCC,
// the current SCC may simply need to be skipped if invalid.
if (UR.InvalidatedSCCs.count(C)) {
LLVM_DEBUG(dbgs() << "Skipping invalidated root or island SCC!\n");
break;
}
// Check that we didn't miss any update scenario.
assert(C->begin() != C->end() && "Cannot have an empty SCC!");
// We handle invalidating the CGSCC analysis manager's information
// for the (potentially updated) SCC here. Note that any other SCCs
// whose structure has changed should have been invalidated by
// whatever was updating the call graph. This SCC gets invalidated
// late as it contains the nodes that were actively being
// processed.
CGAM.invalidate(*C, PassPA);
// Then intersect the preserved set so that invalidation of module
// analyses will eventually occur when the module pass completes.
// Also intersect with the cross-SCC preserved set to capture any
// cross-SCC invalidation.
UR.CrossSCCPA.intersect(PassPA);
PA.intersect(std::move(PassPA));
// The pass may have restructured the call graph and refined the
// current SCC and/or RefSCC. We need to update our current SCC and
// RefSCC pointers to follow these. Also, when the current SCC is
// refined, re-run the SCC pass over the newly refined SCC in order
// to observe the most precise SCC model available. This inherently
// cannot cycle excessively as it only happens when we split SCCs
// apart, at most converging on a DAG of single nodes.
// FIXME: If we ever start having RefSCC passes, we'll want to
// iterate there too.
if (UR.UpdatedC)
LLVM_DEBUG(dbgs()
<< "Re-running SCC passes after a refinement of the "
"current SCC: "
<< *UR.UpdatedC << "\n");
// Note that both `C` and `RC` may at this point refer to deleted,
// invalid SCC and RefSCCs respectively. But we will short circuit
// the processing when we check them in the loop above.
} while (UR.UpdatedC);
} while (!CWorklist.empty());
// We only need to keep internal inlined edge information within
// a RefSCC, clear it to save on space and let the next time we visit
// any of these functions have a fresh start.
InlinedInternalEdges.clear();
} while (!RCWorklist.empty());
}
// By definition we preserve the call garph, all SCC analyses, and the
// analysis proxies by handling them above and in any nested pass managers.
PA.preserveSet<AllAnalysesOn<LazyCallGraph::SCC>>();
PA.preserve<LazyCallGraphAnalysis>();
PA.preserve<CGSCCAnalysisManagerModuleProxy>();
PA.preserve<FunctionAnalysisManagerModuleProxy>();
return PA;
}
PreservedAnalyses DevirtSCCRepeatedPass::run(LazyCallGraph::SCC &InitialC,
CGSCCAnalysisManager &AM,
LazyCallGraph &CG,
CGSCCUpdateResult &UR) {
PreservedAnalyses PA = PreservedAnalyses::all();
PassInstrumentation PI =
AM.getResult<PassInstrumentationAnalysis>(InitialC, CG);
// The SCC may be refined while we are running passes over it, so set up
// a pointer that we can update.
LazyCallGraph::SCC *C = &InitialC;
// Struct to track the counts of direct and indirect calls in each function
// of the SCC.
struct CallCount {
int Direct;
int Indirect;
};
// Put value handles on all of the indirect calls and return the number of
// direct calls for each function in the SCC.
auto ScanSCC = [](LazyCallGraph::SCC &C,
SmallMapVector<Value *, WeakTrackingVH, 16> &CallHandles) {
assert(CallHandles.empty() && "Must start with a clear set of handles.");
SmallDenseMap<Function *, CallCount> CallCounts;
CallCount CountLocal = {0, 0};
for (LazyCallGraph::Node &N : C) {
CallCount &Count =
CallCounts.insert(std::make_pair(&N.getFunction(), CountLocal))
.first->second;
for (Instruction &I : instructions(N.getFunction()))
if (auto *CB = dyn_cast<CallBase>(&I)) {
if (CB->getCalledFunction()) {
++Count.Direct;
} else {
++Count.Indirect;
CallHandles.insert({CB, WeakTrackingVH(CB)});
}
}
}
return CallCounts;
};
UR.IndirectVHs.clear();
// Populate the initial call handles and get the initial call counts.
auto CallCounts = ScanSCC(*C, UR.IndirectVHs);
for (int Iteration = 0;; ++Iteration) {
if (!PI.runBeforePass<LazyCallGraph::SCC>(*Pass, *C))
continue;
PreservedAnalyses PassPA = Pass->run(*C, AM, CG, UR);
if (UR.InvalidatedSCCs.count(C))
PI.runAfterPassInvalidated<LazyCallGraph::SCC>(*Pass, PassPA);
else
PI.runAfterPass<LazyCallGraph::SCC>(*Pass, *C, PassPA);
// If the SCC structure has changed, bail immediately and let the outer
// CGSCC layer handle any iteration to reflect the refined structure.
if (UR.UpdatedC && UR.UpdatedC != C) {
PA.intersect(std::move(PassPA));
break;
}
// Check that we didn't miss any update scenario.
assert(!UR.InvalidatedSCCs.count(C) && "Processing an invalid SCC!");
assert(C->begin() != C->end() && "Cannot have an empty SCC!");
// Check whether any of the handles were devirtualized.
bool Devirt = llvm::any_of(UR.IndirectVHs, [](auto &P) -> bool {
if (P.second) {
if (CallBase *CB = dyn_cast<CallBase>(P.second)) {
if (CB->getCalledFunction()) {
LLVM_DEBUG(dbgs() << "Found devirtualized call: " << *CB << "\n");
return true;
}
}
}
return false;
});
// Rescan to build up a new set of handles and count how many direct
// calls remain. If we decide to iterate, this also sets up the input to
// the next iteration.
UR.IndirectVHs.clear();
auto NewCallCounts = ScanSCC(*C, UR.IndirectVHs);
// If we haven't found an explicit devirtualization already see if we
// have decreased the number of indirect calls and increased the number
// of direct calls for any function in the SCC. This can be fooled by all
// manner of transformations such as DCE and other things, but seems to
// work well in practice.
if (!Devirt)
// Iterate over the keys in NewCallCounts, if Function also exists in
// CallCounts, make the check below.
for (auto &Pair : NewCallCounts) {
auto &CallCountNew = Pair.second;
auto CountIt = CallCounts.find(Pair.first);
if (CountIt != CallCounts.end()) {
const auto &CallCountOld = CountIt->second;
if (CallCountOld.Indirect > CallCountNew.Indirect &&
CallCountOld.Direct < CallCountNew.Direct) {
Devirt = true;
break;
}
}
}
if (!Devirt) {
PA.intersect(std::move(PassPA));
break;
}
// Otherwise, if we've already hit our max, we're done.
if (Iteration >= MaxIterations) {
maxDevirtIterationsReached();
LLVM_DEBUG(
dbgs() << "Found another devirtualization after hitting the max "
"number of repetitions ("
<< MaxIterations << ") on SCC: " << *C << "\n");
PA.intersect(std::move(PassPA));
break;
}
LLVM_DEBUG(
dbgs() << "Repeating an SCC pass after finding a devirtualization in: "
<< *C << "\n");
// Move over the new call counts in preparation for iterating.
CallCounts = std::move(NewCallCounts);
// Update the analysis manager with each run and intersect the total set
// of preserved analyses so we're ready to iterate.
AM.invalidate(*C, PassPA);
PA.intersect(std::move(PassPA));
}
// Note that we don't add any preserved entries here unlike a more normal
// "pass manager" because we only handle invalidation *between* iterations,
// not after the last iteration.
return PA;
}
PreservedAnalyses CGSCCToFunctionPassAdaptor::run(LazyCallGraph::SCC &C,
CGSCCAnalysisManager &AM,
LazyCallGraph &CG,
CGSCCUpdateResult &UR) {
// Setup the function analysis manager from its proxy.
FunctionAnalysisManager &FAM =
AM.getResult<FunctionAnalysisManagerCGSCCProxy>(C, CG).getManager();
SmallVector<LazyCallGraph::Node *, 4> Nodes;
for (LazyCallGraph::Node &N : C)
Nodes.push_back(&N);
// The SCC may get split while we are optimizing functions due to deleting
// edges. If this happens, the current SCC can shift, so keep track of
// a pointer we can overwrite.
LazyCallGraph::SCC *CurrentC = &C;
LLVM_DEBUG(dbgs() << "Running function passes across an SCC: " << C << "\n");
PreservedAnalyses PA = PreservedAnalyses::all();
for (LazyCallGraph::Node *N : Nodes) {
// Skip nodes from other SCCs. These may have been split out during
// processing. We'll eventually visit those SCCs and pick up the nodes
// there.
if (CG.lookupSCC(*N) != CurrentC)
continue;
Function &F = N->getFunction();
PassInstrumentation PI = FAM.getResult<PassInstrumentationAnalysis>(F);
if (!PI.runBeforePass<Function>(*Pass, F))
continue;
PreservedAnalyses PassPA;
{
TimeTraceScope TimeScope(Pass->name());
PassPA = Pass->run(F, FAM);
}
PI.runAfterPass<Function>(*Pass, F, PassPA);
// We know that the function pass couldn't have invalidated any other
// function's analyses (that's the contract of a function pass), so
// directly handle the function analysis manager's invalidation here.
FAM.invalidate(F, PassPA);
// Then intersect the preserved set so that invalidation of module
// analyses will eventually occur when the module pass completes.
PA.intersect(std::move(PassPA));
// If the call graph hasn't been preserved, update it based on this
// function pass. This may also update the current SCC to point to
// a smaller, more refined SCC.
auto PAC = PA.getChecker<LazyCallGraphAnalysis>();
if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Module>>()) {
CurrentC = &updateCGAndAnalysisManagerForFunctionPass(CG, *CurrentC, *N,
AM, UR, FAM);
assert(CG.lookupSCC(*N) == CurrentC &&
"Current SCC not updated to the SCC containing the current node!");
}
}
// By definition we preserve the proxy. And we preserve all analyses on
// Functions. This precludes *any* invalidation of function analyses by the
// proxy, but that's OK because we've taken care to invalidate analyses in
// the function analysis manager incrementally above.
PA.preserveSet<AllAnalysesOn<Function>>();
PA.preserve<FunctionAnalysisManagerCGSCCProxy>();
// We've also ensured that we updated the call graph along the way.
PA.preserve<LazyCallGraphAnalysis>();
return PA;
}
[PM] Support invalidation of inner analysis managers from a pass over the outer IR unit. Summary: This never really got implemented, and was very hard to test before a lot of the refactoring changes to make things more robust. But now we can test it thoroughly and cleanly, especially at the CGSCC level. The core idea is that when an inner analysis manager proxy receives the invalidation event for the outer IR unit, it needs to walk the inner IR units and propagate it to the inner analysis manager for each of those units. For example, each function in the SCC needs to get an invalidation event when the SCC gets one. The function / module interaction is somewhat boring here. This really becomes interesting in the face of analysis-backed IR units. This patch effectively handles all of the CGSCC layer's needs -- both invalidating SCC analysis and invalidating function analysis when an SCC gets invalidated. However, this second aspect doesn't really handle the LoopAnalysisManager well at this point. That one will need some change of design in order to fully integrate, because unlike the call graph, the entire function behind a LoopAnalysis's results can vanish out from under us, and we won't even have a cached API to access. I'd like to try to separate solving the loop problems into a subsequent patch though in order to keep this more focused so I've adapted them to the API and updated the tests that immediately fail, but I've not added the level of testing and validation at that layer that I have at the CGSCC layer. An important aspect of this change is that the proxy for the FunctionAnalysisManager at the SCC pass layer doesn't work like the other proxies for an inner IR unit as it doesn't directly manage the FunctionAnalysisManager and invalidation or clearing of it. This would create an ever worsening problem of dual ownership of this responsibility, split between the module-level FAM proxy and this SCC-level FAM proxy. Instead, this patch changes the SCC-level FAM proxy to work in terms of the module-level proxy and defer to it to handle much of the updates. It only does SCC-specific invalidation. This will become more important in subsequent patches that support more complex invalidaiton scenarios. Reviewers: jlebar Subscribers: mehdi_amini, mcrosier, mzolotukhin, llvm-commits Differential Revision: https://reviews.llvm.org/D27197 llvm-svn: 289317
2016-12-10 14:34:44 +08:00
bool CGSCCAnalysisManagerModuleProxy::Result::invalidate(
Module &M, const PreservedAnalyses &PA,
ModuleAnalysisManager::Invalidator &Inv) {
[PM] Introduce the facilities for registering cross-IR-unit dependencies that require deferred invalidation. This handles the other real-world invalidation scenario that we have cases of: a function analysis which caches references to a module analysis. We currently do this in the AA aggregation layer and might well do this in other places as well. Since this is relative rare, the technique is somewhat more cumbersome. Analyses need to register themselves when accessing the outer analysis manager's proxy. This proxy is already necessarily present to allow access to the outer IR unit's analyses. By registering here we can track and trigger invalidation when that outer analysis goes away. To make this work we need to enhance the PreservedAnalyses infrastructure to support a (slightly) more explicit model for "sets" of analyses, and allow abandoning a single specific analyses even when a set covering that analysis is preserved. That allows us to describe the scenario of preserving all Function analyses *except* for the one where deferred invalidation has triggered. We also need to teach the invalidator API to support direct ID calls instead of always going through a template to dispatch so that we can just record the ID mapping. I've introduced testing of all of this both for simple module<->function cases as well as for more complex cases involving a CGSCC layer. Much like the previous patch I've not tried to fully update the loop pass management layer because that layer is due to be heavily reworked to use similar techniques to the CGSCC to handle updates. As that happens, we'll have a better testing basis for adding support like this. Many thanks to both Justin and Sean for the extensive reviews on this to help bring the API design and documentation into a better state. Differential Revision: https://reviews.llvm.org/D27198 llvm-svn: 290594
2016-12-27 16:40:39 +08:00
// If literally everything is preserved, we're done.
if (PA.areAllPreserved())
return false; // This is still a valid proxy.
[PM] Support invalidation of inner analysis managers from a pass over the outer IR unit. Summary: This never really got implemented, and was very hard to test before a lot of the refactoring changes to make things more robust. But now we can test it thoroughly and cleanly, especially at the CGSCC level. The core idea is that when an inner analysis manager proxy receives the invalidation event for the outer IR unit, it needs to walk the inner IR units and propagate it to the inner analysis manager for each of those units. For example, each function in the SCC needs to get an invalidation event when the SCC gets one. The function / module interaction is somewhat boring here. This really becomes interesting in the face of analysis-backed IR units. This patch effectively handles all of the CGSCC layer's needs -- both invalidating SCC analysis and invalidating function analysis when an SCC gets invalidated. However, this second aspect doesn't really handle the LoopAnalysisManager well at this point. That one will need some change of design in order to fully integrate, because unlike the call graph, the entire function behind a LoopAnalysis's results can vanish out from under us, and we won't even have a cached API to access. I'd like to try to separate solving the loop problems into a subsequent patch though in order to keep this more focused so I've adapted them to the API and updated the tests that immediately fail, but I've not added the level of testing and validation at that layer that I have at the CGSCC layer. An important aspect of this change is that the proxy for the FunctionAnalysisManager at the SCC pass layer doesn't work like the other proxies for an inner IR unit as it doesn't directly manage the FunctionAnalysisManager and invalidation or clearing of it. This would create an ever worsening problem of dual ownership of this responsibility, split between the module-level FAM proxy and this SCC-level FAM proxy. Instead, this patch changes the SCC-level FAM proxy to work in terms of the module-level proxy and defer to it to handle much of the updates. It only does SCC-specific invalidation. This will become more important in subsequent patches that support more complex invalidaiton scenarios. Reviewers: jlebar Subscribers: mehdi_amini, mcrosier, mzolotukhin, llvm-commits Differential Revision: https://reviews.llvm.org/D27197 llvm-svn: 289317
2016-12-10 14:34:44 +08:00
// If this proxy or the call graph is going to be invalidated, we also need
// to clear all the keys coming from that analysis.
//
// We also directly invalidate the FAM's module proxy if necessary, and if
// that proxy isn't preserved we can't preserve this proxy either. We rely on
// it to handle module -> function analysis invalidation in the face of
// structural changes and so if it's unavailable we conservatively clear the
[PM] Introduce the facilities for registering cross-IR-unit dependencies that require deferred invalidation. This handles the other real-world invalidation scenario that we have cases of: a function analysis which caches references to a module analysis. We currently do this in the AA aggregation layer and might well do this in other places as well. Since this is relative rare, the technique is somewhat more cumbersome. Analyses need to register themselves when accessing the outer analysis manager's proxy. This proxy is already necessarily present to allow access to the outer IR unit's analyses. By registering here we can track and trigger invalidation when that outer analysis goes away. To make this work we need to enhance the PreservedAnalyses infrastructure to support a (slightly) more explicit model for "sets" of analyses, and allow abandoning a single specific analyses even when a set covering that analysis is preserved. That allows us to describe the scenario of preserving all Function analyses *except* for the one where deferred invalidation has triggered. We also need to teach the invalidator API to support direct ID calls instead of always going through a template to dispatch so that we can just record the ID mapping. I've introduced testing of all of this both for simple module<->function cases as well as for more complex cases involving a CGSCC layer. Much like the previous patch I've not tried to fully update the loop pass management layer because that layer is due to be heavily reworked to use similar techniques to the CGSCC to handle updates. As that happens, we'll have a better testing basis for adding support like this. Many thanks to both Justin and Sean for the extensive reviews on this to help bring the API design and documentation into a better state. Differential Revision: https://reviews.llvm.org/D27198 llvm-svn: 290594
2016-12-27 16:40:39 +08:00
// entire SCC layer as well rather than trying to do invalidation ourselves.
auto PAC = PA.getChecker<CGSCCAnalysisManagerModuleProxy>();
if (!(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Module>>()) ||
[PM] Support invalidation of inner analysis managers from a pass over the outer IR unit. Summary: This never really got implemented, and was very hard to test before a lot of the refactoring changes to make things more robust. But now we can test it thoroughly and cleanly, especially at the CGSCC level. The core idea is that when an inner analysis manager proxy receives the invalidation event for the outer IR unit, it needs to walk the inner IR units and propagate it to the inner analysis manager for each of those units. For example, each function in the SCC needs to get an invalidation event when the SCC gets one. The function / module interaction is somewhat boring here. This really becomes interesting in the face of analysis-backed IR units. This patch effectively handles all of the CGSCC layer's needs -- both invalidating SCC analysis and invalidating function analysis when an SCC gets invalidated. However, this second aspect doesn't really handle the LoopAnalysisManager well at this point. That one will need some change of design in order to fully integrate, because unlike the call graph, the entire function behind a LoopAnalysis's results can vanish out from under us, and we won't even have a cached API to access. I'd like to try to separate solving the loop problems into a subsequent patch though in order to keep this more focused so I've adapted them to the API and updated the tests that immediately fail, but I've not added the level of testing and validation at that layer that I have at the CGSCC layer. An important aspect of this change is that the proxy for the FunctionAnalysisManager at the SCC pass layer doesn't work like the other proxies for an inner IR unit as it doesn't directly manage the FunctionAnalysisManager and invalidation or clearing of it. This would create an ever worsening problem of dual ownership of this responsibility, split between the module-level FAM proxy and this SCC-level FAM proxy. Instead, this patch changes the SCC-level FAM proxy to work in terms of the module-level proxy and defer to it to handle much of the updates. It only does SCC-specific invalidation. This will become more important in subsequent patches that support more complex invalidaiton scenarios. Reviewers: jlebar Subscribers: mehdi_amini, mcrosier, mzolotukhin, llvm-commits Differential Revision: https://reviews.llvm.org/D27197 llvm-svn: 289317
2016-12-10 14:34:44 +08:00
Inv.invalidate<LazyCallGraphAnalysis>(M, PA) ||
Inv.invalidate<FunctionAnalysisManagerModuleProxy>(M, PA)) {
InnerAM->clear();
// And the proxy itself should be marked as invalid so that we can observe
// the new call graph. This isn't strictly necessary because we cheat
// above, but is still useful.
return true;
}
[PM] Introduce the facilities for registering cross-IR-unit dependencies that require deferred invalidation. This handles the other real-world invalidation scenario that we have cases of: a function analysis which caches references to a module analysis. We currently do this in the AA aggregation layer and might well do this in other places as well. Since this is relative rare, the technique is somewhat more cumbersome. Analyses need to register themselves when accessing the outer analysis manager's proxy. This proxy is already necessarily present to allow access to the outer IR unit's analyses. By registering here we can track and trigger invalidation when that outer analysis goes away. To make this work we need to enhance the PreservedAnalyses infrastructure to support a (slightly) more explicit model for "sets" of analyses, and allow abandoning a single specific analyses even when a set covering that analysis is preserved. That allows us to describe the scenario of preserving all Function analyses *except* for the one where deferred invalidation has triggered. We also need to teach the invalidator API to support direct ID calls instead of always going through a template to dispatch so that we can just record the ID mapping. I've introduced testing of all of this both for simple module<->function cases as well as for more complex cases involving a CGSCC layer. Much like the previous patch I've not tried to fully update the loop pass management layer because that layer is due to be heavily reworked to use similar techniques to the CGSCC to handle updates. As that happens, we'll have a better testing basis for adding support like this. Many thanks to both Justin and Sean for the extensive reviews on this to help bring the API design and documentation into a better state. Differential Revision: https://reviews.llvm.org/D27198 llvm-svn: 290594
2016-12-27 16:40:39 +08:00
// Directly check if the relevant set is preserved so we can short circuit
// invalidating SCCs below.
bool AreSCCAnalysesPreserved =
PA.allAnalysesInSetPreserved<AllAnalysesOn<LazyCallGraph::SCC>>();
[PM] Support invalidation of inner analysis managers from a pass over the outer IR unit. Summary: This never really got implemented, and was very hard to test before a lot of the refactoring changes to make things more robust. But now we can test it thoroughly and cleanly, especially at the CGSCC level. The core idea is that when an inner analysis manager proxy receives the invalidation event for the outer IR unit, it needs to walk the inner IR units and propagate it to the inner analysis manager for each of those units. For example, each function in the SCC needs to get an invalidation event when the SCC gets one. The function / module interaction is somewhat boring here. This really becomes interesting in the face of analysis-backed IR units. This patch effectively handles all of the CGSCC layer's needs -- both invalidating SCC analysis and invalidating function analysis when an SCC gets invalidated. However, this second aspect doesn't really handle the LoopAnalysisManager well at this point. That one will need some change of design in order to fully integrate, because unlike the call graph, the entire function behind a LoopAnalysis's results can vanish out from under us, and we won't even have a cached API to access. I'd like to try to separate solving the loop problems into a subsequent patch though in order to keep this more focused so I've adapted them to the API and updated the tests that immediately fail, but I've not added the level of testing and validation at that layer that I have at the CGSCC layer. An important aspect of this change is that the proxy for the FunctionAnalysisManager at the SCC pass layer doesn't work like the other proxies for an inner IR unit as it doesn't directly manage the FunctionAnalysisManager and invalidation or clearing of it. This would create an ever worsening problem of dual ownership of this responsibility, split between the module-level FAM proxy and this SCC-level FAM proxy. Instead, this patch changes the SCC-level FAM proxy to work in terms of the module-level proxy and defer to it to handle much of the updates. It only does SCC-specific invalidation. This will become more important in subsequent patches that support more complex invalidaiton scenarios. Reviewers: jlebar Subscribers: mehdi_amini, mcrosier, mzolotukhin, llvm-commits Differential Revision: https://reviews.llvm.org/D27197 llvm-svn: 289317
2016-12-10 14:34:44 +08:00
// Ok, we have a graph, so we can propagate the invalidation down into it.
[PM/LCG] Remove the lazy RefSCC formation from the LazyCallGraph during iteration. The lazy formation of RefSCCs isn't really the most important part of the laziness here -- that has to do with walking the functions themselves -- and isn't essential to maintain. Originally, there were incremental update algorithms that relied on updates happening predominantly near the most recent RefSCC formed, but those have been replaced with ones that have much tighter general case bounds at this point. We do still perform asserts that only scale well due to this incrementality, but those are easy to place behind EXPENSIVE_CHECKS. Removing this simplifies the entire analysis by having a single up-front step that builds all of the RefSCCs in a direct Tarjan walk. We can even easily replace this with other or better algorithms at will and with much less confusion now that there is no iterator-based incremental logic involved. This removes a lot of complexity from LCG. Another advantage of moving in this direction is that it simplifies testing the system substantially as we no longer have to worry about observing and mutating the graph half-way through the RefSCC formation. We still need a somewhat special iterator for RefSCCs because we want the iterator to remain stable in the face of graph updates. However, this now merely involves relative indexing to the current RefSCC's position in the sequence which isn't too hard. Differential Revision: https://reviews.llvm.org/D29381 llvm-svn: 294227
2017-02-07 03:38:06 +08:00
G->buildRefSCCs();
[PM] Support invalidation of inner analysis managers from a pass over the outer IR unit. Summary: This never really got implemented, and was very hard to test before a lot of the refactoring changes to make things more robust. But now we can test it thoroughly and cleanly, especially at the CGSCC level. The core idea is that when an inner analysis manager proxy receives the invalidation event for the outer IR unit, it needs to walk the inner IR units and propagate it to the inner analysis manager for each of those units. For example, each function in the SCC needs to get an invalidation event when the SCC gets one. The function / module interaction is somewhat boring here. This really becomes interesting in the face of analysis-backed IR units. This patch effectively handles all of the CGSCC layer's needs -- both invalidating SCC analysis and invalidating function analysis when an SCC gets invalidated. However, this second aspect doesn't really handle the LoopAnalysisManager well at this point. That one will need some change of design in order to fully integrate, because unlike the call graph, the entire function behind a LoopAnalysis's results can vanish out from under us, and we won't even have a cached API to access. I'd like to try to separate solving the loop problems into a subsequent patch though in order to keep this more focused so I've adapted them to the API and updated the tests that immediately fail, but I've not added the level of testing and validation at that layer that I have at the CGSCC layer. An important aspect of this change is that the proxy for the FunctionAnalysisManager at the SCC pass layer doesn't work like the other proxies for an inner IR unit as it doesn't directly manage the FunctionAnalysisManager and invalidation or clearing of it. This would create an ever worsening problem of dual ownership of this responsibility, split between the module-level FAM proxy and this SCC-level FAM proxy. Instead, this patch changes the SCC-level FAM proxy to work in terms of the module-level proxy and defer to it to handle much of the updates. It only does SCC-specific invalidation. This will become more important in subsequent patches that support more complex invalidaiton scenarios. Reviewers: jlebar Subscribers: mehdi_amini, mcrosier, mzolotukhin, llvm-commits Differential Revision: https://reviews.llvm.org/D27197 llvm-svn: 289317
2016-12-10 14:34:44 +08:00
for (auto &RC : G->postorder_ref_sccs())
[PM] Introduce the facilities for registering cross-IR-unit dependencies that require deferred invalidation. This handles the other real-world invalidation scenario that we have cases of: a function analysis which caches references to a module analysis. We currently do this in the AA aggregation layer and might well do this in other places as well. Since this is relative rare, the technique is somewhat more cumbersome. Analyses need to register themselves when accessing the outer analysis manager's proxy. This proxy is already necessarily present to allow access to the outer IR unit's analyses. By registering here we can track and trigger invalidation when that outer analysis goes away. To make this work we need to enhance the PreservedAnalyses infrastructure to support a (slightly) more explicit model for "sets" of analyses, and allow abandoning a single specific analyses even when a set covering that analysis is preserved. That allows us to describe the scenario of preserving all Function analyses *except* for the one where deferred invalidation has triggered. We also need to teach the invalidator API to support direct ID calls instead of always going through a template to dispatch so that we can just record the ID mapping. I've introduced testing of all of this both for simple module<->function cases as well as for more complex cases involving a CGSCC layer. Much like the previous patch I've not tried to fully update the loop pass management layer because that layer is due to be heavily reworked to use similar techniques to the CGSCC to handle updates. As that happens, we'll have a better testing basis for adding support like this. Many thanks to both Justin and Sean for the extensive reviews on this to help bring the API design and documentation into a better state. Differential Revision: https://reviews.llvm.org/D27198 llvm-svn: 290594
2016-12-27 16:40:39 +08:00
for (auto &C : RC) {
Optional<PreservedAnalyses> InnerPA;
// Check to see whether the preserved set needs to be adjusted based on
// module-level analysis invalidation triggering deferred invalidation
// for this SCC.
if (auto *OuterProxy =
InnerAM->getCachedResult<ModuleAnalysisManagerCGSCCProxy>(C))
for (const auto &OuterInvalidationPair :
OuterProxy->getOuterInvalidations()) {
AnalysisKey *OuterAnalysisID = OuterInvalidationPair.first;
const auto &InnerAnalysisIDs = OuterInvalidationPair.second;
if (Inv.invalidate(OuterAnalysisID, M, PA)) {
if (!InnerPA)
InnerPA = PA;
for (AnalysisKey *InnerAnalysisID : InnerAnalysisIDs)
InnerPA->abandon(InnerAnalysisID);
}
}
// Check if we needed a custom PA set. If so we'll need to run the inner
// invalidation.
if (InnerPA) {
InnerAM->invalidate(C, *InnerPA);
continue;
}
// Otherwise we only need to do invalidation if the original PA set didn't
// preserve all SCC analyses.
if (!AreSCCAnalysesPreserved)
InnerAM->invalidate(C, PA);
}
[PM] Support invalidation of inner analysis managers from a pass over the outer IR unit. Summary: This never really got implemented, and was very hard to test before a lot of the refactoring changes to make things more robust. But now we can test it thoroughly and cleanly, especially at the CGSCC level. The core idea is that when an inner analysis manager proxy receives the invalidation event for the outer IR unit, it needs to walk the inner IR units and propagate it to the inner analysis manager for each of those units. For example, each function in the SCC needs to get an invalidation event when the SCC gets one. The function / module interaction is somewhat boring here. This really becomes interesting in the face of analysis-backed IR units. This patch effectively handles all of the CGSCC layer's needs -- both invalidating SCC analysis and invalidating function analysis when an SCC gets invalidated. However, this second aspect doesn't really handle the LoopAnalysisManager well at this point. That one will need some change of design in order to fully integrate, because unlike the call graph, the entire function behind a LoopAnalysis's results can vanish out from under us, and we won't even have a cached API to access. I'd like to try to separate solving the loop problems into a subsequent patch though in order to keep this more focused so I've adapted them to the API and updated the tests that immediately fail, but I've not added the level of testing and validation at that layer that I have at the CGSCC layer. An important aspect of this change is that the proxy for the FunctionAnalysisManager at the SCC pass layer doesn't work like the other proxies for an inner IR unit as it doesn't directly manage the FunctionAnalysisManager and invalidation or clearing of it. This would create an ever worsening problem of dual ownership of this responsibility, split between the module-level FAM proxy and this SCC-level FAM proxy. Instead, this patch changes the SCC-level FAM proxy to work in terms of the module-level proxy and defer to it to handle much of the updates. It only does SCC-specific invalidation. This will become more important in subsequent patches that support more complex invalidaiton scenarios. Reviewers: jlebar Subscribers: mehdi_amini, mcrosier, mzolotukhin, llvm-commits Differential Revision: https://reviews.llvm.org/D27197 llvm-svn: 289317
2016-12-10 14:34:44 +08:00
// Return false to indicate that this result is still a valid proxy.
return false;
}
template <>
CGSCCAnalysisManagerModuleProxy::Result
CGSCCAnalysisManagerModuleProxy::run(Module &M, ModuleAnalysisManager &AM) {
// Force the Function analysis manager to also be available so that it can
// be accessed in an SCC analysis and proxied onward to function passes.
// FIXME: It is pretty awkward to just drop the result here and assert that
// we can find it again later.
(void)AM.getResult<FunctionAnalysisManagerModuleProxy>(M);
return Result(*InnerAM, AM.getResult<LazyCallGraphAnalysis>(M));
}
AnalysisKey FunctionAnalysisManagerCGSCCProxy::Key;
FunctionAnalysisManagerCGSCCProxy::Result
FunctionAnalysisManagerCGSCCProxy::run(LazyCallGraph::SCC &C,
CGSCCAnalysisManager &AM,
LazyCallGraph &CG) {
// Note: unconditionally getting checking that the proxy exists may get it at
// this point. There are cases when this is being run unnecessarily, but
// it is cheap and having the assertion in place is more valuable.
auto &MAMProxy = AM.getResult<ModuleAnalysisManagerCGSCCProxy>(C, CG);
[PM] Support invalidation of inner analysis managers from a pass over the outer IR unit. Summary: This never really got implemented, and was very hard to test before a lot of the refactoring changes to make things more robust. But now we can test it thoroughly and cleanly, especially at the CGSCC level. The core idea is that when an inner analysis manager proxy receives the invalidation event for the outer IR unit, it needs to walk the inner IR units and propagate it to the inner analysis manager for each of those units. For example, each function in the SCC needs to get an invalidation event when the SCC gets one. The function / module interaction is somewhat boring here. This really becomes interesting in the face of analysis-backed IR units. This patch effectively handles all of the CGSCC layer's needs -- both invalidating SCC analysis and invalidating function analysis when an SCC gets invalidated. However, this second aspect doesn't really handle the LoopAnalysisManager well at this point. That one will need some change of design in order to fully integrate, because unlike the call graph, the entire function behind a LoopAnalysis's results can vanish out from under us, and we won't even have a cached API to access. I'd like to try to separate solving the loop problems into a subsequent patch though in order to keep this more focused so I've adapted them to the API and updated the tests that immediately fail, but I've not added the level of testing and validation at that layer that I have at the CGSCC layer. An important aspect of this change is that the proxy for the FunctionAnalysisManager at the SCC pass layer doesn't work like the other proxies for an inner IR unit as it doesn't directly manage the FunctionAnalysisManager and invalidation or clearing of it. This would create an ever worsening problem of dual ownership of this responsibility, split between the module-level FAM proxy and this SCC-level FAM proxy. Instead, this patch changes the SCC-level FAM proxy to work in terms of the module-level proxy and defer to it to handle much of the updates. It only does SCC-specific invalidation. This will become more important in subsequent patches that support more complex invalidaiton scenarios. Reviewers: jlebar Subscribers: mehdi_amini, mcrosier, mzolotukhin, llvm-commits Differential Revision: https://reviews.llvm.org/D27197 llvm-svn: 289317
2016-12-10 14:34:44 +08:00
Module &M = *C.begin()->getFunction().getParent();
bool ProxyExists =
MAMProxy.cachedResultExists<FunctionAnalysisManagerModuleProxy>(M);
assert(ProxyExists &&
"The CGSCC pass manager requires that the FAM module proxy is run "
"on the module prior to entering the CGSCC walk");
(void)ProxyExists;
// We just return an empty result. The caller will use the updateFAM interface
// to correctly register the relevant FunctionAnalysisManager based on the
// context in which this proxy is run.
return Result();
[PM] Support invalidation of inner analysis managers from a pass over the outer IR unit. Summary: This never really got implemented, and was very hard to test before a lot of the refactoring changes to make things more robust. But now we can test it thoroughly and cleanly, especially at the CGSCC level. The core idea is that when an inner analysis manager proxy receives the invalidation event for the outer IR unit, it needs to walk the inner IR units and propagate it to the inner analysis manager for each of those units. For example, each function in the SCC needs to get an invalidation event when the SCC gets one. The function / module interaction is somewhat boring here. This really becomes interesting in the face of analysis-backed IR units. This patch effectively handles all of the CGSCC layer's needs -- both invalidating SCC analysis and invalidating function analysis when an SCC gets invalidated. However, this second aspect doesn't really handle the LoopAnalysisManager well at this point. That one will need some change of design in order to fully integrate, because unlike the call graph, the entire function behind a LoopAnalysis's results can vanish out from under us, and we won't even have a cached API to access. I'd like to try to separate solving the loop problems into a subsequent patch though in order to keep this more focused so I've adapted them to the API and updated the tests that immediately fail, but I've not added the level of testing and validation at that layer that I have at the CGSCC layer. An important aspect of this change is that the proxy for the FunctionAnalysisManager at the SCC pass layer doesn't work like the other proxies for an inner IR unit as it doesn't directly manage the FunctionAnalysisManager and invalidation or clearing of it. This would create an ever worsening problem of dual ownership of this responsibility, split between the module-level FAM proxy and this SCC-level FAM proxy. Instead, this patch changes the SCC-level FAM proxy to work in terms of the module-level proxy and defer to it to handle much of the updates. It only does SCC-specific invalidation. This will become more important in subsequent patches that support more complex invalidaiton scenarios. Reviewers: jlebar Subscribers: mehdi_amini, mcrosier, mzolotukhin, llvm-commits Differential Revision: https://reviews.llvm.org/D27197 llvm-svn: 289317
2016-12-10 14:34:44 +08:00
}
bool FunctionAnalysisManagerCGSCCProxy::Result::invalidate(
LazyCallGraph::SCC &C, const PreservedAnalyses &PA,
CGSCCAnalysisManager::Invalidator &Inv) {
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
// If literally everything is preserved, we're done.
if (PA.areAllPreserved())
return false; // This is still a valid proxy.
// All updates to preserve valid results are done below, so we don't need to
// invalidate this proxy.
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
//
// Note that in order to preserve this proxy, a module pass must ensure that
// the FAM has been completely updated to handle the deletion of functions.
// Specifically, any FAM-cached results for those functions need to have been
// forcibly cleared. When preserved, this proxy will only invalidate results
// cached on functions *still in the module* at the end of the module pass.
auto PAC = PA.getChecker<FunctionAnalysisManagerCGSCCProxy>();
if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<LazyCallGraph::SCC>>()) {
for (LazyCallGraph::Node &N : C)
FAM->clear(N.getFunction(), N.getFunction().getName());
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
return false;
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
}
// Directly check if the relevant set is preserved.
bool AreFunctionAnalysesPreserved =
PA.allAnalysesInSetPreserved<AllAnalysesOn<Function>>();
// Now walk all the functions to see if any inner analysis invalidation is
// necessary.
for (LazyCallGraph::Node &N : C) {
Function &F = N.getFunction();
Optional<PreservedAnalyses> FunctionPA;
// Check to see whether the preserved set needs to be pruned based on
// SCC-level analysis invalidation that triggers deferred invalidation
// registered with the outer analysis manager proxy for this function.
if (auto *OuterProxy =
FAM->getCachedResult<CGSCCAnalysisManagerFunctionProxy>(F))
for (const auto &OuterInvalidationPair :
OuterProxy->getOuterInvalidations()) {
AnalysisKey *OuterAnalysisID = OuterInvalidationPair.first;
const auto &InnerAnalysisIDs = OuterInvalidationPair.second;
if (Inv.invalidate(OuterAnalysisID, C, PA)) {
if (!FunctionPA)
FunctionPA = PA;
for (AnalysisKey *InnerAnalysisID : InnerAnalysisIDs)
FunctionPA->abandon(InnerAnalysisID);
}
}
// Check if we needed a custom PA set, and if so we'll need to run the
// inner invalidation.
if (FunctionPA) {
FAM->invalidate(F, *FunctionPA);
continue;
}
// Otherwise we only need to do invalidation if the original PA set didn't
// preserve all function analyses.
if (!AreFunctionAnalysesPreserved)
FAM->invalidate(F, PA);
}
[PM] Support invalidation of inner analysis managers from a pass over the outer IR unit. Summary: This never really got implemented, and was very hard to test before a lot of the refactoring changes to make things more robust. But now we can test it thoroughly and cleanly, especially at the CGSCC level. The core idea is that when an inner analysis manager proxy receives the invalidation event for the outer IR unit, it needs to walk the inner IR units and propagate it to the inner analysis manager for each of those units. For example, each function in the SCC needs to get an invalidation event when the SCC gets one. The function / module interaction is somewhat boring here. This really becomes interesting in the face of analysis-backed IR units. This patch effectively handles all of the CGSCC layer's needs -- both invalidating SCC analysis and invalidating function analysis when an SCC gets invalidated. However, this second aspect doesn't really handle the LoopAnalysisManager well at this point. That one will need some change of design in order to fully integrate, because unlike the call graph, the entire function behind a LoopAnalysis's results can vanish out from under us, and we won't even have a cached API to access. I'd like to try to separate solving the loop problems into a subsequent patch though in order to keep this more focused so I've adapted them to the API and updated the tests that immediately fail, but I've not added the level of testing and validation at that layer that I have at the CGSCC layer. An important aspect of this change is that the proxy for the FunctionAnalysisManager at the SCC pass layer doesn't work like the other proxies for an inner IR unit as it doesn't directly manage the FunctionAnalysisManager and invalidation or clearing of it. This would create an ever worsening problem of dual ownership of this responsibility, split between the module-level FAM proxy and this SCC-level FAM proxy. Instead, this patch changes the SCC-level FAM proxy to work in terms of the module-level proxy and defer to it to handle much of the updates. It only does SCC-specific invalidation. This will become more important in subsequent patches that support more complex invalidaiton scenarios. Reviewers: jlebar Subscribers: mehdi_amini, mcrosier, mzolotukhin, llvm-commits Differential Revision: https://reviews.llvm.org/D27197 llvm-svn: 289317
2016-12-10 14:34:44 +08:00
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
// Return false to indicate that this result is still a valid proxy.
[PM] Support invalidation of inner analysis managers from a pass over the outer IR unit. Summary: This never really got implemented, and was very hard to test before a lot of the refactoring changes to make things more robust. But now we can test it thoroughly and cleanly, especially at the CGSCC level. The core idea is that when an inner analysis manager proxy receives the invalidation event for the outer IR unit, it needs to walk the inner IR units and propagate it to the inner analysis manager for each of those units. For example, each function in the SCC needs to get an invalidation event when the SCC gets one. The function / module interaction is somewhat boring here. This really becomes interesting in the face of analysis-backed IR units. This patch effectively handles all of the CGSCC layer's needs -- both invalidating SCC analysis and invalidating function analysis when an SCC gets invalidated. However, this second aspect doesn't really handle the LoopAnalysisManager well at this point. That one will need some change of design in order to fully integrate, because unlike the call graph, the entire function behind a LoopAnalysis's results can vanish out from under us, and we won't even have a cached API to access. I'd like to try to separate solving the loop problems into a subsequent patch though in order to keep this more focused so I've adapted them to the API and updated the tests that immediately fail, but I've not added the level of testing and validation at that layer that I have at the CGSCC layer. An important aspect of this change is that the proxy for the FunctionAnalysisManager at the SCC pass layer doesn't work like the other proxies for an inner IR unit as it doesn't directly manage the FunctionAnalysisManager and invalidation or clearing of it. This would create an ever worsening problem of dual ownership of this responsibility, split between the module-level FAM proxy and this SCC-level FAM proxy. Instead, this patch changes the SCC-level FAM proxy to work in terms of the module-level proxy and defer to it to handle much of the updates. It only does SCC-specific invalidation. This will become more important in subsequent patches that support more complex invalidaiton scenarios. Reviewers: jlebar Subscribers: mehdi_amini, mcrosier, mzolotukhin, llvm-commits Differential Revision: https://reviews.llvm.org/D27197 llvm-svn: 289317
2016-12-10 14:34:44 +08:00
return false;
}
} // end namespace llvm
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
/// When a new SCC is created for the graph we first update the
/// FunctionAnalysisManager in the Proxy's result.
/// As there might be function analysis results cached for the functions now in
/// that SCC, two forms of updates are required.
///
/// First, a proxy from the SCC to the FunctionAnalysisManager needs to be
/// created so that any subsequent invalidation events to the SCC are
/// propagated to the function analysis results cached for functions within it.
///
/// Second, if any of the functions within the SCC have analysis results with
/// outer analysis dependencies, then those dependencies would point to the
/// *wrong* SCC's analysis result. We forcibly invalidate the necessary
/// function analyses so that they don't retain stale handles.
static void updateNewSCCFunctionAnalyses(LazyCallGraph::SCC &C,
LazyCallGraph &G,
CGSCCAnalysisManager &AM,
FunctionAnalysisManager &FAM) {
AM.getResult<FunctionAnalysisManagerCGSCCProxy>(C, G).updateFAM(FAM);
// Now walk the functions in this SCC and invalidate any function analysis
// results that might have outer dependencies on an SCC analysis.
for (LazyCallGraph::Node &N : C) {
Function &F = N.getFunction();
auto *OuterProxy =
FAM.getCachedResult<CGSCCAnalysisManagerFunctionProxy>(F);
if (!OuterProxy)
// No outer analyses were queried, nothing to do.
continue;
// Forcibly abandon all the inner analyses with dependencies, but
// invalidate nothing else.
auto PA = PreservedAnalyses::all();
for (const auto &OuterInvalidationPair :
OuterProxy->getOuterInvalidations()) {
const auto &InnerAnalysisIDs = OuterInvalidationPair.second;
for (AnalysisKey *InnerAnalysisID : InnerAnalysisIDs)
PA.abandon(InnerAnalysisID);
}
// Now invalidate anything we found.
FAM.invalidate(F, PA);
}
}
void llvm::maxDevirtIterationsReached() {
if (AbortOnMaxDevirtIterationsReached)
report_fatal_error("Max devirtualization iterations reached");
}
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
/// Helper function to update both the \c CGSCCAnalysisManager \p AM and the \c
/// CGSCCPassManager's \c CGSCCUpdateResult \p UR based on a range of newly
/// added SCCs.
///
/// The range of new SCCs must be in postorder already. The SCC they were split
/// out of must be provided as \p C. The current node being mutated and
/// triggering updates must be passed as \p N.
///
/// This function returns the SCC containing \p N. This will be either \p C if
/// no new SCCs have been split out, or it will be the new SCC containing \p N.
template <typename SCCRangeT>
static LazyCallGraph::SCC *
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
incorporateNewSCCRange(const SCCRangeT &NewSCCRange, LazyCallGraph &G,
LazyCallGraph::Node &N, LazyCallGraph::SCC *C,
CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR) {
using SCC = LazyCallGraph::SCC;
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
if (NewSCCRange.begin() == NewSCCRange.end())
return C;
[PM] Teach the CGSCC's CG update utility to more carefully invalidate analyses when we're about to break apart an SCC. We can't wait until after breaking apart the SCC to invalidate things: 1) Which SCC do we then invalidate? All of them? 2) Even if we invalidate all of them, a newly created SCC may not have a proxy that will convey the invalidation to functions! Previously we only invalidated one of the SCCs and too late. This led to stale analyses remaining in the cache. And because the caching strategy actually works, they would get used and chaos would ensue. Doing invalidation early is somewhat pessimizing though if we *know* that the SCC structure won't change. So it turns out that the design to make the mutation API force the caller to know the *kind* of mutation in advance was indeed 100% correct and we didn't do enough of it. So this change also splits two cases of switching a call edge to a ref edge into two separate APIs so that callers can clearly test for this and take the easy path without invalidating when appropriate. This is particularly important in this case as we expect most inlines to be between functions in separate SCCs and so the common case is that we don't have to so aggressively invalidate analyses. The LCG API change in turn needed some basic cleanups and better testing in its unittest. No interesting functionality changed there other than more coverage of the returned sequence of SCCs. While this seems like an obvious improvement over the current state, I'd like to revisit the core concept of invalidating within the CG-update layer at all. I'm wondering if we would be better served forcing the callers to handle the invalidation beforehand in the cases that they can handle it. An interesting example is when we want to teach the inliner to *update and preserve* analyses. But we can cross that bridge when we get there. With this patch, the new pass manager an build all of the LLVM test suite at -O3 and everything passes. =D I haven't bootstrapped yet and I'm sure there are still plenty of bugs, but this gives a nice baseline so I'm going to increasingly focus on fleshing out the missing functionality, especially the bits that are just turned off right now in order to let us establish this baseline. llvm-svn: 290664
2016-12-28 18:34:50 +08:00
// Add the current SCC to the worklist as its shape has changed.
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
UR.CWorklist.insert(C);
LLVM_DEBUG(dbgs() << "Enqueuing the existing SCC in the worklist:" << *C
<< "\n");
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
SCC *OldC = C;
// Update the current SCC. Note that if we have new SCCs, this must actually
// change the SCC.
assert(C != &*NewSCCRange.begin() &&
"Cannot insert new SCCs without changing current SCC!");
C = &*NewSCCRange.begin();
assert(G.lookupSCC(N) == C && "Failed to update current SCC!");
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
// If we had a cached FAM proxy originally, we will want to create more of
// them for each SCC that was split off.
FunctionAnalysisManager *FAM = nullptr;
if (auto *FAMProxy =
AM.getCachedResult<FunctionAnalysisManagerCGSCCProxy>(*OldC))
FAM = &FAMProxy->getManager();
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
// We need to propagate an invalidation call to all but the newly current SCC
// because the outer pass manager won't do that for us after splitting them.
// FIXME: We should accept a PreservedAnalysis from the CG updater so that if
// there are preserved analysis we can avoid invalidating them here for
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
// split-off SCCs.
// We know however that this will preserve any FAM proxy so go ahead and mark
// that.
PreservedAnalyses PA;
PA.preserve<FunctionAnalysisManagerCGSCCProxy>();
AM.invalidate(*OldC, PA);
// Ensure the now-current SCC's function analyses are updated.
if (FAM)
updateNewSCCFunctionAnalyses(*C, G, AM, *FAM);
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
for (SCC &NewC : llvm::reverse(make_range(std::next(NewSCCRange.begin()),
NewSCCRange.end()))) {
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
assert(C != &NewC && "No need to re-visit the current SCC!");
assert(OldC != &NewC && "Already handled the original SCC!");
UR.CWorklist.insert(&NewC);
LLVM_DEBUG(dbgs() << "Enqueuing a newly formed SCC:" << NewC << "\n");
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
// Ensure new SCCs' function analyses are updated.
if (FAM)
updateNewSCCFunctionAnalyses(NewC, G, AM, *FAM);
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
// Also propagate a normal invalidation to the new SCC as only the current
// will get one from the pass manager infrastructure.
[PM] Finish implementing and fix a chain of bugs uncovered by testing the invalidation propagation logic from an SCC to a Function. I wrote the infrastructure to test this but didn't actually use it in the unit test where it was designed to be used. =[ My bad. Once I actually added it to the test case I discovered that it also hadn't been properly implemented, so I've implemented it. The logic in the FAM proxy for an SCC pass to propagate invalidation follows the same ideas as the FAM proxy for a Module pass, but the implementation is a bit different to reflect the fact that it is forwarding just for an SCC. However, implementing this correctly uncovered a surprising "bug" (it was conservatively correct but relatively very expensive) in how we handle invalidation when splitting one SCC into multiple SCCs. We did an eager invalidation when in reality we should be deferring invaliadtion for the *current* SCC to the CGSCC pass manager and just invaliating the newly constructed SCCs. Otherwise we end up invalidating too much too soon. This was exposed by the inliner test case that I've updated. Now, we invalidate *just* the split off '(test1_f)' SCC when doing the CG update, and then the inliner finishes and invalidates the '(test1_g, test1_h)' SCC's analyses. The first few attempts at fixing this hit still more bugs, but all of those are covered by existing tests. For example, the inliner should also preserve the FAM proxy to avoid unnecesasry invalidation, and this is safe because the CG update routines it uses handle any necessary adjustments to the FAM proxy. Finally, the unittests for the CGSCC pass manager needed a bunch of updates where we weren't correctly preserving the FAM proxy because it hadn't been fully implemented and failing to preserve it didn't matter. Note that this doesn't yet fix the current crasher due to MemSSA finding a stale dominator tree, but without this the fix to that crasher doesn't really make any sense when testing because it relies on the proxy behavior. llvm-svn: 307487
2017-07-09 11:59:31 +08:00
AM.invalidate(NewC, PA);
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
}
return C;
}
static LazyCallGraph::SCC &updateCGAndAnalysisManagerForPass(
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
LazyCallGraph &G, LazyCallGraph::SCC &InitialC, LazyCallGraph::Node &N,
CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR,
FunctionAnalysisManager &FAM, bool FunctionPass) {
using Node = LazyCallGraph::Node;
using Edge = LazyCallGraph::Edge;
using SCC = LazyCallGraph::SCC;
using RefSCC = LazyCallGraph::RefSCC;
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
RefSCC &InitialRC = InitialC.getOuterRefSCC();
SCC *C = &InitialC;
RefSCC *RC = &InitialRC;
Function &F = N.getFunction();
// Walk the function body and build up the set of retained, promoted, and
// demoted edges.
SmallVector<Constant *, 16> Worklist;
SmallPtrSet<Constant *, 16> Visited;
[PM/LCG] Teach the LazyCallGraph how to replace a function without disturbing the graph or having to update edges. This is motivated by porting argument promotion to the new pass manager. Because of how LLVM IR Function objects work, in order to change their signature a new object needs to be created. This is efficient and straight forward in the IR but previously was very hard to implement in LCG. We could easily replace the function a node in the graph represents. The challenging part is how to handle updating the edges in the graph. LCG previously used an edge to a raw function to represent a node that had not yet been scanned for calls and references. This was the core of its laziness. However, that model causes this kind of update to be very hard: 1) The keys to lookup an edge need to be `Function*`s that would all need to be updated when we update the node. 2) There will be some unknown number of edges that haven't transitioned from `Function*` edges to `Node*` edges. All of this complexity isn't necessary. Instead, we can always build a node around any function, always pointing edges at it and always using it as the key to lookup an edge. To maintain the laziness, we need to sink the *edges* of a node into a secondary object and explicitly model transitioning a node from empty to populated by scanning the function. This design seems much cleaner in a number of ways, but importantly there is now exactly *one* place where the `Function*` has to be updated! Some other cleanups that fall out of this include having something to model the *entry* edges more accurately. Rather than hand rolling parts of the node in the graph itself, we have an explicit `EdgeSequence` object that gives us exactly the functionality needed. We also have a consistent place to define the edge iterators and can use them for both the entry edges and the internal edges of the graph. The API used to model the separation between a node and its edges is intentionally very thin as most clients are expected to deal with nodes that have populated edges. We model this exactly as an optional does with an additional method to populate the edges when that is a reasonable thing for a client to do. This is based on API design suggestions from Richard Smith and David Blaikie, credit goes to them for helping pick how to model this without it being either too explicit or too implicit. The patch is somewhat noisy due to shifting around iterator types and new syntax for walking the edges of a node, but most of the functionality change is in the `Edge`, `EdgeSequence`, and `Node` types. Differential Revision: https://reviews.llvm.org/D29577 llvm-svn: 294653
2017-02-10 07:24:13 +08:00
SmallPtrSet<Node *, 16> RetainedEdges;
SmallSetVector<Node *, 4> PromotedRefTargets;
SmallSetVector<Node *, 4> DemotedCallTargets;
SmallSetVector<Node *, 4> NewCallEdges;
SmallSetVector<Node *, 4> NewRefEdges;
SmallSetVector<Node *, 4> NewNodes;
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// First walk the function and handle all called functions. We do this first
// because if there is a single call edge, whether there are ref edges is
// irrelevant.
Reland [CGSCC] Detect devirtualization in more cases The devirtualization wrapper misses cases where if it wraps a pass manager, an individual pass may devirtualize an indirect call created by a previous pass. For example, inlining may create a new indirect call which is devirtualized by instcombine. Currently the devirtualization wrapper will not see that because it only checks cgscc edges at the very beginning and end of the pass (manager) it wraps. This fixes some tests testing this exact behavior in the legacy PM. Instead of checking WeakTrackingVHs for CallBases at the very beginning and end of the pass it wraps, check every time updateCGAndAnalysisManagerForPass() is called. check-llvm and check-clang with -abort-on-max-devirt-iterations-reached on by default doesn't show any failures outside of tests specifically testing it so it doesn't needlessly rerun passes more than necessary. (The NPM -O2/3 pipeline run the inliner/function simplification pipeline under a devirtualization repeater pass up to 4 times by default). http://llvm-compile-time-tracker.com/?config=O3&stat=instructions&remote=aeubanks shows that 7zip has ~1% compile time regression. I looked at it and saw that there indeed was devirtualization happening that was not previously caught, so now it reruns the CGSCC pipeline on some SCCs, which is WAI. The initial land assumed CallBase WeakTrackingVHs would always be CallBases, but they can be RAUW'd with undef. Reviewed By: asbirlea Differential Revision: https://reviews.llvm.org/D89587
2020-10-15 08:56:38 +08:00
for (Instruction &I : instructions(F)) {
if (auto *CB = dyn_cast<CallBase>(&I)) {
if (Function *Callee = CB->getCalledFunction()) {
if (Visited.insert(Callee).second && !Callee->isDeclaration()) {
Node *CalleeN = G.lookup(*Callee);
if (!CalleeN) {
CalleeN = &G.get(*Callee);
NewNodes.insert(CalleeN);
}
Edge *E = N->lookup(*CalleeN);
assert((E || !FunctionPass) &&
"No function transformations should introduce *new* "
"call edges! Any new calls should be modeled as "
"promoted existing ref edges!");
bool Inserted = RetainedEdges.insert(CalleeN).second;
(void)Inserted;
assert(Inserted && "We should never visit a function twice.");
if (!E)
NewCallEdges.insert(CalleeN);
else if (!E->isCall())
PromotedRefTargets.insert(CalleeN);
}
Reland [CGSCC] Detect devirtualization in more cases The devirtualization wrapper misses cases where if it wraps a pass manager, an individual pass may devirtualize an indirect call created by a previous pass. For example, inlining may create a new indirect call which is devirtualized by instcombine. Currently the devirtualization wrapper will not see that because it only checks cgscc edges at the very beginning and end of the pass (manager) it wraps. This fixes some tests testing this exact behavior in the legacy PM. Instead of checking WeakTrackingVHs for CallBases at the very beginning and end of the pass it wraps, check every time updateCGAndAnalysisManagerForPass() is called. check-llvm and check-clang with -abort-on-max-devirt-iterations-reached on by default doesn't show any failures outside of tests specifically testing it so it doesn't needlessly rerun passes more than necessary. (The NPM -O2/3 pipeline run the inliner/function simplification pipeline under a devirtualization repeater pass up to 4 times by default). http://llvm-compile-time-tracker.com/?config=O3&stat=instructions&remote=aeubanks shows that 7zip has ~1% compile time regression. I looked at it and saw that there indeed was devirtualization happening that was not previously caught, so now it reruns the CGSCC pipeline on some SCCs, which is WAI. The initial land assumed CallBase WeakTrackingVHs would always be CallBases, but they can be RAUW'd with undef. Reviewed By: asbirlea Differential Revision: https://reviews.llvm.org/D89587
2020-10-15 08:56:38 +08:00
} else {
// We can miss devirtualization if an indirect call is created then
// promoted before updateCGAndAnalysisManagerForPass runs.
auto *Entry = UR.IndirectVHs.find(CB);
if (Entry == UR.IndirectVHs.end())
UR.IndirectVHs.insert({CB, WeakTrackingVH(CB)});
else if (!Entry->second)
Entry->second = WeakTrackingVH(CB);
}
}
}
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// Now walk all references.
for (Instruction &I : instructions(F))
for (Value *Op : I.operand_values())
if (auto *OpC = dyn_cast<Constant>(Op))
if (Visited.insert(OpC).second)
Worklist.push_back(OpC);
auto VisitRef = [&](Function &Referee) {
Node *RefereeN = G.lookup(Referee);
if (!RefereeN) {
RefereeN = &G.get(Referee);
NewNodes.insert(RefereeN);
}
Edge *E = N->lookup(*RefereeN);
assert((E || !FunctionPass) &&
"No function transformations should introduce *new* ref "
"edges! Any new ref edges would require IPO which "
"function passes aren't allowed to do!");
bool Inserted = RetainedEdges.insert(RefereeN).second;
(void)Inserted;
assert(Inserted && "We should never visit a function twice.");
if (!E)
NewRefEdges.insert(RefereeN);
else if (E->isCall())
DemotedCallTargets.insert(RefereeN);
};
LazyCallGraph::visitReferences(Worklist, Visited, VisitRef);
for (Node *NewNode : NewNodes)
G.initNode(*NewNode, *C);
// Handle new ref edges.
for (Node *RefTarget : NewRefEdges) {
SCC &TargetC = *G.lookupSCC(*RefTarget);
RefSCC &TargetRC = TargetC.getOuterRefSCC();
(void)TargetRC;
// TODO: This only allows trivial edges to be added for now.
assert((RC == &TargetRC ||
RC->isAncestorOf(TargetRC)) && "New ref edge is not trivial!");
RC->insertTrivialRefEdge(N, *RefTarget);
}
// Handle new call edges.
for (Node *CallTarget : NewCallEdges) {
SCC &TargetC = *G.lookupSCC(*CallTarget);
RefSCC &TargetRC = TargetC.getOuterRefSCC();
(void)TargetRC;
// TODO: This only allows trivial edges to be added for now.
assert((RC == &TargetRC ||
RC->isAncestorOf(TargetRC)) && "New call edge is not trivial!");
// Add a trivial ref edge to be promoted later on alongside
// PromotedRefTargets.
RC->insertTrivialRefEdge(N, *CallTarget);
}
// Include synthetic reference edges to known, defined lib functions.
for (auto *LibFn : G.getLibFunctions())
// While the list of lib functions doesn't have repeats, don't re-visit
// anything handled above.
if (!Visited.count(LibFn))
VisitRef(*LibFn);
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// First remove all of the edges that are no longer present in this function.
[LCG] Switch one of the update methods for the LazyCallGraph to support limited batch updates. Specifically, allow removing multiple reference edges starting from a common source node. There are a few constraints that play into supporting this form of batching: 1) The way updates occur during the CGSCC walk, about the most we can functionally batch together are those with a common source node. This also makes the batching simpler to implement, so it seems a worthwhile restriction. 2) The far and away hottest function for large C++ files I measured (generated code for protocol buffers) showed a huge amount of time was spent removing ref edges specifically, so it seems worth focusing there. 3) The algorithm for removing ref edges is very amenable to this restricted batching. There are just both API and implementation special casing for the non-batch case that gets in the way. Once removed, supporting batches is nearly trivial. This does modify the API in an interesting way -- now, we only preserve the target RefSCC when the RefSCC structure is unchanged. In the face of any splits, we create brand new RefSCC objects. However, all of the users were OK with it that I could find. Only the unittest needed interesting updates here. How much does batching these updates help? I instrumented the compiler when run over a very large generated source file for a protocol buffer and found that the majority of updates are intrinsically updating one function at a time. However, nearly 40% of the total ref edges removed are removed as part of a batch of removals greater than one, so these are the cases batching can help with. When compiling the IR for this file with 'opt' and 'O3', this patch reduces the total time by 8-9%. Differential Revision: https://reviews.llvm.org/D36352 llvm-svn: 310450
2017-08-09 17:05:27 +08:00
// The first step makes these edges uniformly ref edges and accumulates them
// into a separate data structure so removal doesn't invalidate anything.
SmallVector<Node *, 4> DeadTargets;
for (Edge &E : *N) {
if (RetainedEdges.count(&E.getNode()))
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
continue;
[LCG] Switch one of the update methods for the LazyCallGraph to support limited batch updates. Specifically, allow removing multiple reference edges starting from a common source node. There are a few constraints that play into supporting this form of batching: 1) The way updates occur during the CGSCC walk, about the most we can functionally batch together are those with a common source node. This also makes the batching simpler to implement, so it seems a worthwhile restriction. 2) The far and away hottest function for large C++ files I measured (generated code for protocol buffers) showed a huge amount of time was spent removing ref edges specifically, so it seems worth focusing there. 3) The algorithm for removing ref edges is very amenable to this restricted batching. There are just both API and implementation special casing for the non-batch case that gets in the way. Once removed, supporting batches is nearly trivial. This does modify the API in an interesting way -- now, we only preserve the target RefSCC when the RefSCC structure is unchanged. In the face of any splits, we create brand new RefSCC objects. However, all of the users were OK with it that I could find. Only the unittest needed interesting updates here. How much does batching these updates help? I instrumented the compiler when run over a very large generated source file for a protocol buffer and found that the majority of updates are intrinsically updating one function at a time. However, nearly 40% of the total ref edges removed are removed as part of a batch of removals greater than one, so these are the cases batching can help with. When compiling the IR for this file with 'opt' and 'O3', this patch reduces the total time by 8-9%. Differential Revision: https://reviews.llvm.org/D36352 llvm-svn: 310450
2017-08-09 17:05:27 +08:00
SCC &TargetC = *G.lookupSCC(E.getNode());
RefSCC &TargetRC = TargetC.getOuterRefSCC();
if (&TargetRC == RC && E.isCall()) {
[PM] Teach the CGSCC's CG update utility to more carefully invalidate analyses when we're about to break apart an SCC. We can't wait until after breaking apart the SCC to invalidate things: 1) Which SCC do we then invalidate? All of them? 2) Even if we invalidate all of them, a newly created SCC may not have a proxy that will convey the invalidation to functions! Previously we only invalidated one of the SCCs and too late. This led to stale analyses remaining in the cache. And because the caching strategy actually works, they would get used and chaos would ensue. Doing invalidation early is somewhat pessimizing though if we *know* that the SCC structure won't change. So it turns out that the design to make the mutation API force the caller to know the *kind* of mutation in advance was indeed 100% correct and we didn't do enough of it. So this change also splits two cases of switching a call edge to a ref edge into two separate APIs so that callers can clearly test for this and take the easy path without invalidating when appropriate. This is particularly important in this case as we expect most inlines to be between functions in separate SCCs and so the common case is that we don't have to so aggressively invalidate analyses. The LCG API change in turn needed some basic cleanups and better testing in its unittest. No interesting functionality changed there other than more coverage of the returned sequence of SCCs. While this seems like an obvious improvement over the current state, I'd like to revisit the core concept of invalidating within the CG-update layer at all. I'm wondering if we would be better served forcing the callers to handle the invalidation beforehand in the cases that they can handle it. An interesting example is when we want to teach the inliner to *update and preserve* analyses. But we can cross that bridge when we get there. With this patch, the new pass manager an build all of the LLVM test suite at -O3 and everything passes. =D I haven't bootstrapped yet and I'm sure there are still plenty of bugs, but this gives a nice baseline so I'm going to increasingly focus on fleshing out the missing functionality, especially the bits that are just turned off right now in order to let us establish this baseline. llvm-svn: 290664
2016-12-28 18:34:50 +08:00
if (C != &TargetC) {
// For separate SCCs this is trivial.
[LCG] Switch one of the update methods for the LazyCallGraph to support limited batch updates. Specifically, allow removing multiple reference edges starting from a common source node. There are a few constraints that play into supporting this form of batching: 1) The way updates occur during the CGSCC walk, about the most we can functionally batch together are those with a common source node. This also makes the batching simpler to implement, so it seems a worthwhile restriction. 2) The far and away hottest function for large C++ files I measured (generated code for protocol buffers) showed a huge amount of time was spent removing ref edges specifically, so it seems worth focusing there. 3) The algorithm for removing ref edges is very amenable to this restricted batching. There are just both API and implementation special casing for the non-batch case that gets in the way. Once removed, supporting batches is nearly trivial. This does modify the API in an interesting way -- now, we only preserve the target RefSCC when the RefSCC structure is unchanged. In the face of any splits, we create brand new RefSCC objects. However, all of the users were OK with it that I could find. Only the unittest needed interesting updates here. How much does batching these updates help? I instrumented the compiler when run over a very large generated source file for a protocol buffer and found that the majority of updates are intrinsically updating one function at a time. However, nearly 40% of the total ref edges removed are removed as part of a batch of removals greater than one, so these are the cases batching can help with. When compiling the IR for this file with 'opt' and 'O3', this patch reduces the total time by 8-9%. Differential Revision: https://reviews.llvm.org/D36352 llvm-svn: 310450
2017-08-09 17:05:27 +08:00
RC->switchTrivialInternalEdgeToRef(N, E.getNode());
[PM] Teach the CGSCC's CG update utility to more carefully invalidate analyses when we're about to break apart an SCC. We can't wait until after breaking apart the SCC to invalidate things: 1) Which SCC do we then invalidate? All of them? 2) Even if we invalidate all of them, a newly created SCC may not have a proxy that will convey the invalidation to functions! Previously we only invalidated one of the SCCs and too late. This led to stale analyses remaining in the cache. And because the caching strategy actually works, they would get used and chaos would ensue. Doing invalidation early is somewhat pessimizing though if we *know* that the SCC structure won't change. So it turns out that the design to make the mutation API force the caller to know the *kind* of mutation in advance was indeed 100% correct and we didn't do enough of it. So this change also splits two cases of switching a call edge to a ref edge into two separate APIs so that callers can clearly test for this and take the easy path without invalidating when appropriate. This is particularly important in this case as we expect most inlines to be between functions in separate SCCs and so the common case is that we don't have to so aggressively invalidate analyses. The LCG API change in turn needed some basic cleanups and better testing in its unittest. No interesting functionality changed there other than more coverage of the returned sequence of SCCs. While this seems like an obvious improvement over the current state, I'd like to revisit the core concept of invalidating within the CG-update layer at all. I'm wondering if we would be better served forcing the callers to handle the invalidation beforehand in the cases that they can handle it. An interesting example is when we want to teach the inliner to *update and preserve* analyses. But we can cross that bridge when we get there. With this patch, the new pass manager an build all of the LLVM test suite at -O3 and everything passes. =D I haven't bootstrapped yet and I'm sure there are still plenty of bugs, but this gives a nice baseline so I'm going to increasingly focus on fleshing out the missing functionality, especially the bits that are just turned off right now in order to let us establish this baseline. llvm-svn: 290664
2016-12-28 18:34:50 +08:00
} else {
// Now update the call graph.
[LCG] Switch one of the update methods for the LazyCallGraph to support limited batch updates. Specifically, allow removing multiple reference edges starting from a common source node. There are a few constraints that play into supporting this form of batching: 1) The way updates occur during the CGSCC walk, about the most we can functionally batch together are those with a common source node. This also makes the batching simpler to implement, so it seems a worthwhile restriction. 2) The far and away hottest function for large C++ files I measured (generated code for protocol buffers) showed a huge amount of time was spent removing ref edges specifically, so it seems worth focusing there. 3) The algorithm for removing ref edges is very amenable to this restricted batching. There are just both API and implementation special casing for the non-batch case that gets in the way. Once removed, supporting batches is nearly trivial. This does modify the API in an interesting way -- now, we only preserve the target RefSCC when the RefSCC structure is unchanged. In the face of any splits, we create brand new RefSCC objects. However, all of the users were OK with it that I could find. Only the unittest needed interesting updates here. How much does batching these updates help? I instrumented the compiler when run over a very large generated source file for a protocol buffer and found that the majority of updates are intrinsically updating one function at a time. However, nearly 40% of the total ref edges removed are removed as part of a batch of removals greater than one, so these are the cases batching can help with. When compiling the IR for this file with 'opt' and 'O3', this patch reduces the total time by 8-9%. Differential Revision: https://reviews.llvm.org/D36352 llvm-svn: 310450
2017-08-09 17:05:27 +08:00
C = incorporateNewSCCRange(RC->switchInternalEdgeToRef(N, E.getNode()),
G, N, C, AM, UR);
[PM] Teach the CGSCC's CG update utility to more carefully invalidate analyses when we're about to break apart an SCC. We can't wait until after breaking apart the SCC to invalidate things: 1) Which SCC do we then invalidate? All of them? 2) Even if we invalidate all of them, a newly created SCC may not have a proxy that will convey the invalidation to functions! Previously we only invalidated one of the SCCs and too late. This led to stale analyses remaining in the cache. And because the caching strategy actually works, they would get used and chaos would ensue. Doing invalidation early is somewhat pessimizing though if we *know* that the SCC structure won't change. So it turns out that the design to make the mutation API force the caller to know the *kind* of mutation in advance was indeed 100% correct and we didn't do enough of it. So this change also splits two cases of switching a call edge to a ref edge into two separate APIs so that callers can clearly test for this and take the easy path without invalidating when appropriate. This is particularly important in this case as we expect most inlines to be between functions in separate SCCs and so the common case is that we don't have to so aggressively invalidate analyses. The LCG API change in turn needed some basic cleanups and better testing in its unittest. No interesting functionality changed there other than more coverage of the returned sequence of SCCs. While this seems like an obvious improvement over the current state, I'd like to revisit the core concept of invalidating within the CG-update layer at all. I'm wondering if we would be better served forcing the callers to handle the invalidation beforehand in the cases that they can handle it. An interesting example is when we want to teach the inliner to *update and preserve* analyses. But we can cross that bridge when we get there. With this patch, the new pass manager an build all of the LLVM test suite at -O3 and everything passes. =D I haven't bootstrapped yet and I'm sure there are still plenty of bugs, but this gives a nice baseline so I'm going to increasingly focus on fleshing out the missing functionality, especially the bits that are just turned off right now in order to let us establish this baseline. llvm-svn: 290664
2016-12-28 18:34:50 +08:00
}
}
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
[LCG] Switch one of the update methods for the LazyCallGraph to support limited batch updates. Specifically, allow removing multiple reference edges starting from a common source node. There are a few constraints that play into supporting this form of batching: 1) The way updates occur during the CGSCC walk, about the most we can functionally batch together are those with a common source node. This also makes the batching simpler to implement, so it seems a worthwhile restriction. 2) The far and away hottest function for large C++ files I measured (generated code for protocol buffers) showed a huge amount of time was spent removing ref edges specifically, so it seems worth focusing there. 3) The algorithm for removing ref edges is very amenable to this restricted batching. There are just both API and implementation special casing for the non-batch case that gets in the way. Once removed, supporting batches is nearly trivial. This does modify the API in an interesting way -- now, we only preserve the target RefSCC when the RefSCC structure is unchanged. In the face of any splits, we create brand new RefSCC objects. However, all of the users were OK with it that I could find. Only the unittest needed interesting updates here. How much does batching these updates help? I instrumented the compiler when run over a very large generated source file for a protocol buffer and found that the majority of updates are intrinsically updating one function at a time. However, nearly 40% of the total ref edges removed are removed as part of a batch of removals greater than one, so these are the cases batching can help with. When compiling the IR for this file with 'opt' and 'O3', this patch reduces the total time by 8-9%. Differential Revision: https://reviews.llvm.org/D36352 llvm-svn: 310450
2017-08-09 17:05:27 +08:00
// Now that this is ready for actual removal, put it into our list.
DeadTargets.push_back(&E.getNode());
}
// Remove the easy cases quickly and actually pull them out of our list.
DeadTargets.erase(
llvm::remove_if(DeadTargets,
[&](Node *TargetN) {
SCC &TargetC = *G.lookupSCC(*TargetN);
RefSCC &TargetRC = TargetC.getOuterRefSCC();
// We can't trivially remove internal targets, so skip
// those.
if (&TargetRC == RC)
return false;
RC->removeOutgoingEdge(N, *TargetN);
LLVM_DEBUG(dbgs() << "Deleting outgoing edge from '"
<< N << "' to '" << TargetN << "'\n");
[LCG] Switch one of the update methods for the LazyCallGraph to support limited batch updates. Specifically, allow removing multiple reference edges starting from a common source node. There are a few constraints that play into supporting this form of batching: 1) The way updates occur during the CGSCC walk, about the most we can functionally batch together are those with a common source node. This also makes the batching simpler to implement, so it seems a worthwhile restriction. 2) The far and away hottest function for large C++ files I measured (generated code for protocol buffers) showed a huge amount of time was spent removing ref edges specifically, so it seems worth focusing there. 3) The algorithm for removing ref edges is very amenable to this restricted batching. There are just both API and implementation special casing for the non-batch case that gets in the way. Once removed, supporting batches is nearly trivial. This does modify the API in an interesting way -- now, we only preserve the target RefSCC when the RefSCC structure is unchanged. In the face of any splits, we create brand new RefSCC objects. However, all of the users were OK with it that I could find. Only the unittest needed interesting updates here. How much does batching these updates help? I instrumented the compiler when run over a very large generated source file for a protocol buffer and found that the majority of updates are intrinsically updating one function at a time. However, nearly 40% of the total ref edges removed are removed as part of a batch of removals greater than one, so these are the cases batching can help with. When compiling the IR for this file with 'opt' and 'O3', this patch reduces the total time by 8-9%. Differential Revision: https://reviews.llvm.org/D36352 llvm-svn: 310450
2017-08-09 17:05:27 +08:00
return true;
}),
DeadTargets.end());
// Now do a batch removal of the internal ref edges left.
auto NewRefSCCs = RC->removeInternalRefEdge(N, DeadTargets);
if (!NewRefSCCs.empty()) {
// The old RefSCC is dead, mark it as such.
UR.InvalidatedRefSCCs.insert(RC);
// Note that we don't bother to invalidate analyses as ref-edge
// connectivity is not really observable in any way and is intended
// exclusively to be used for ordering of transforms rather than for
// analysis conclusions.
// Update RC to the "bottom".
assert(G.lookupSCC(N) == C && "Changed the SCC when splitting RefSCCs!");
RC = &C->getOuterRefSCC();
assert(G.lookupRefSCC(N) == RC && "Failed to update current RefSCC!");
// The RC worklist is in reverse postorder, so we enqueue the new ones in
// RPO except for the one which contains the source node as that is the
// "bottom" we will continue processing in the bottom-up walk.
assert(NewRefSCCs.front() == RC &&
"New current RefSCC not first in the returned list!");
for (RefSCC *NewRC : llvm::reverse(make_range(std::next(NewRefSCCs.begin()),
NewRefSCCs.end()))) {
[LCG] Switch one of the update methods for the LazyCallGraph to support limited batch updates. Specifically, allow removing multiple reference edges starting from a common source node. There are a few constraints that play into supporting this form of batching: 1) The way updates occur during the CGSCC walk, about the most we can functionally batch together are those with a common source node. This also makes the batching simpler to implement, so it seems a worthwhile restriction. 2) The far and away hottest function for large C++ files I measured (generated code for protocol buffers) showed a huge amount of time was spent removing ref edges specifically, so it seems worth focusing there. 3) The algorithm for removing ref edges is very amenable to this restricted batching. There are just both API and implementation special casing for the non-batch case that gets in the way. Once removed, supporting batches is nearly trivial. This does modify the API in an interesting way -- now, we only preserve the target RefSCC when the RefSCC structure is unchanged. In the face of any splits, we create brand new RefSCC objects. However, all of the users were OK with it that I could find. Only the unittest needed interesting updates here. How much does batching these updates help? I instrumented the compiler when run over a very large generated source file for a protocol buffer and found that the majority of updates are intrinsically updating one function at a time. However, nearly 40% of the total ref edges removed are removed as part of a batch of removals greater than one, so these are the cases batching can help with. When compiling the IR for this file with 'opt' and 'O3', this patch reduces the total time by 8-9%. Differential Revision: https://reviews.llvm.org/D36352 llvm-svn: 310450
2017-08-09 17:05:27 +08:00
assert(NewRC != RC && "Should not encounter the current RefSCC further "
"in the postorder list of new RefSCCs.");
UR.RCWorklist.insert(NewRC);
LLVM_DEBUG(dbgs() << "Enqueuing a new RefSCC in the update worklist: "
<< *NewRC << "\n");
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
}
}
// Next demote all the call edges that are now ref edges. This helps make
// the SCCs small which should minimize the work below as we don't want to
// form cycles that this would break.
[PM/LCG] Teach the LazyCallGraph how to replace a function without disturbing the graph or having to update edges. This is motivated by porting argument promotion to the new pass manager. Because of how LLVM IR Function objects work, in order to change their signature a new object needs to be created. This is efficient and straight forward in the IR but previously was very hard to implement in LCG. We could easily replace the function a node in the graph represents. The challenging part is how to handle updating the edges in the graph. LCG previously used an edge to a raw function to represent a node that had not yet been scanned for calls and references. This was the core of its laziness. However, that model causes this kind of update to be very hard: 1) The keys to lookup an edge need to be `Function*`s that would all need to be updated when we update the node. 2) There will be some unknown number of edges that haven't transitioned from `Function*` edges to `Node*` edges. All of this complexity isn't necessary. Instead, we can always build a node around any function, always pointing edges at it and always using it as the key to lookup an edge. To maintain the laziness, we need to sink the *edges* of a node into a secondary object and explicitly model transitioning a node from empty to populated by scanning the function. This design seems much cleaner in a number of ways, but importantly there is now exactly *one* place where the `Function*` has to be updated! Some other cleanups that fall out of this include having something to model the *entry* edges more accurately. Rather than hand rolling parts of the node in the graph itself, we have an explicit `EdgeSequence` object that gives us exactly the functionality needed. We also have a consistent place to define the edge iterators and can use them for both the entry edges and the internal edges of the graph. The API used to model the separation between a node and its edges is intentionally very thin as most clients are expected to deal with nodes that have populated edges. We model this exactly as an optional does with an additional method to populate the edges when that is a reasonable thing for a client to do. This is based on API design suggestions from Richard Smith and David Blaikie, credit goes to them for helping pick how to model this without it being either too explicit or too implicit. The patch is somewhat noisy due to shifting around iterator types and new syntax for walking the edges of a node, but most of the functionality change is in the `Edge`, `EdgeSequence`, and `Node` types. Differential Revision: https://reviews.llvm.org/D29577 llvm-svn: 294653
2017-02-10 07:24:13 +08:00
for (Node *RefTarget : DemotedCallTargets) {
SCC &TargetC = *G.lookupSCC(*RefTarget);
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
RefSCC &TargetRC = TargetC.getOuterRefSCC();
// The easy case is when the target RefSCC is not this RefSCC. This is
// only supported when the target RefSCC is a child of this RefSCC.
if (&TargetRC != RC) {
assert(RC->isAncestorOf(TargetRC) &&
"Cannot potentially form RefSCC cycles here!");
[PM/LCG] Teach the LazyCallGraph how to replace a function without disturbing the graph or having to update edges. This is motivated by porting argument promotion to the new pass manager. Because of how LLVM IR Function objects work, in order to change their signature a new object needs to be created. This is efficient and straight forward in the IR but previously was very hard to implement in LCG. We could easily replace the function a node in the graph represents. The challenging part is how to handle updating the edges in the graph. LCG previously used an edge to a raw function to represent a node that had not yet been scanned for calls and references. This was the core of its laziness. However, that model causes this kind of update to be very hard: 1) The keys to lookup an edge need to be `Function*`s that would all need to be updated when we update the node. 2) There will be some unknown number of edges that haven't transitioned from `Function*` edges to `Node*` edges. All of this complexity isn't necessary. Instead, we can always build a node around any function, always pointing edges at it and always using it as the key to lookup an edge. To maintain the laziness, we need to sink the *edges* of a node into a secondary object and explicitly model transitioning a node from empty to populated by scanning the function. This design seems much cleaner in a number of ways, but importantly there is now exactly *one* place where the `Function*` has to be updated! Some other cleanups that fall out of this include having something to model the *entry* edges more accurately. Rather than hand rolling parts of the node in the graph itself, we have an explicit `EdgeSequence` object that gives us exactly the functionality needed. We also have a consistent place to define the edge iterators and can use them for both the entry edges and the internal edges of the graph. The API used to model the separation between a node and its edges is intentionally very thin as most clients are expected to deal with nodes that have populated edges. We model this exactly as an optional does with an additional method to populate the edges when that is a reasonable thing for a client to do. This is based on API design suggestions from Richard Smith and David Blaikie, credit goes to them for helping pick how to model this without it being either too explicit or too implicit. The patch is somewhat noisy due to shifting around iterator types and new syntax for walking the edges of a node, but most of the functionality change is in the `Edge`, `EdgeSequence`, and `Node` types. Differential Revision: https://reviews.llvm.org/D29577 llvm-svn: 294653
2017-02-10 07:24:13 +08:00
RC->switchOutgoingEdgeToRef(N, *RefTarget);
LLVM_DEBUG(dbgs() << "Switch outgoing call edge to a ref edge from '" << N
<< "' to '" << *RefTarget << "'\n");
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
continue;
}
[PM] Teach the CGSCC's CG update utility to more carefully invalidate analyses when we're about to break apart an SCC. We can't wait until after breaking apart the SCC to invalidate things: 1) Which SCC do we then invalidate? All of them? 2) Even if we invalidate all of them, a newly created SCC may not have a proxy that will convey the invalidation to functions! Previously we only invalidated one of the SCCs and too late. This led to stale analyses remaining in the cache. And because the caching strategy actually works, they would get used and chaos would ensue. Doing invalidation early is somewhat pessimizing though if we *know* that the SCC structure won't change. So it turns out that the design to make the mutation API force the caller to know the *kind* of mutation in advance was indeed 100% correct and we didn't do enough of it. So this change also splits two cases of switching a call edge to a ref edge into two separate APIs so that callers can clearly test for this and take the easy path without invalidating when appropriate. This is particularly important in this case as we expect most inlines to be between functions in separate SCCs and so the common case is that we don't have to so aggressively invalidate analyses. The LCG API change in turn needed some basic cleanups and better testing in its unittest. No interesting functionality changed there other than more coverage of the returned sequence of SCCs. While this seems like an obvious improvement over the current state, I'd like to revisit the core concept of invalidating within the CG-update layer at all. I'm wondering if we would be better served forcing the callers to handle the invalidation beforehand in the cases that they can handle it. An interesting example is when we want to teach the inliner to *update and preserve* analyses. But we can cross that bridge when we get there. With this patch, the new pass manager an build all of the LLVM test suite at -O3 and everything passes. =D I haven't bootstrapped yet and I'm sure there are still plenty of bugs, but this gives a nice baseline so I'm going to increasingly focus on fleshing out the missing functionality, especially the bits that are just turned off right now in order to let us establish this baseline. llvm-svn: 290664
2016-12-28 18:34:50 +08:00
// We are switching an internal call edge to a ref edge. This may split up
// some SCCs.
if (C != &TargetC) {
// For separate SCCs this is trivial.
[PM/LCG] Teach the LazyCallGraph how to replace a function without disturbing the graph or having to update edges. This is motivated by porting argument promotion to the new pass manager. Because of how LLVM IR Function objects work, in order to change their signature a new object needs to be created. This is efficient and straight forward in the IR but previously was very hard to implement in LCG. We could easily replace the function a node in the graph represents. The challenging part is how to handle updating the edges in the graph. LCG previously used an edge to a raw function to represent a node that had not yet been scanned for calls and references. This was the core of its laziness. However, that model causes this kind of update to be very hard: 1) The keys to lookup an edge need to be `Function*`s that would all need to be updated when we update the node. 2) There will be some unknown number of edges that haven't transitioned from `Function*` edges to `Node*` edges. All of this complexity isn't necessary. Instead, we can always build a node around any function, always pointing edges at it and always using it as the key to lookup an edge. To maintain the laziness, we need to sink the *edges* of a node into a secondary object and explicitly model transitioning a node from empty to populated by scanning the function. This design seems much cleaner in a number of ways, but importantly there is now exactly *one* place where the `Function*` has to be updated! Some other cleanups that fall out of this include having something to model the *entry* edges more accurately. Rather than hand rolling parts of the node in the graph itself, we have an explicit `EdgeSequence` object that gives us exactly the functionality needed. We also have a consistent place to define the edge iterators and can use them for both the entry edges and the internal edges of the graph. The API used to model the separation between a node and its edges is intentionally very thin as most clients are expected to deal with nodes that have populated edges. We model this exactly as an optional does with an additional method to populate the edges when that is a reasonable thing for a client to do. This is based on API design suggestions from Richard Smith and David Blaikie, credit goes to them for helping pick how to model this without it being either too explicit or too implicit. The patch is somewhat noisy due to shifting around iterator types and new syntax for walking the edges of a node, but most of the functionality change is in the `Edge`, `EdgeSequence`, and `Node` types. Differential Revision: https://reviews.llvm.org/D29577 llvm-svn: 294653
2017-02-10 07:24:13 +08:00
RC->switchTrivialInternalEdgeToRef(N, *RefTarget);
[PM] Teach the CGSCC's CG update utility to more carefully invalidate analyses when we're about to break apart an SCC. We can't wait until after breaking apart the SCC to invalidate things: 1) Which SCC do we then invalidate? All of them? 2) Even if we invalidate all of them, a newly created SCC may not have a proxy that will convey the invalidation to functions! Previously we only invalidated one of the SCCs and too late. This led to stale analyses remaining in the cache. And because the caching strategy actually works, they would get used and chaos would ensue. Doing invalidation early is somewhat pessimizing though if we *know* that the SCC structure won't change. So it turns out that the design to make the mutation API force the caller to know the *kind* of mutation in advance was indeed 100% correct and we didn't do enough of it. So this change also splits two cases of switching a call edge to a ref edge into two separate APIs so that callers can clearly test for this and take the easy path without invalidating when appropriate. This is particularly important in this case as we expect most inlines to be between functions in separate SCCs and so the common case is that we don't have to so aggressively invalidate analyses. The LCG API change in turn needed some basic cleanups and better testing in its unittest. No interesting functionality changed there other than more coverage of the returned sequence of SCCs. While this seems like an obvious improvement over the current state, I'd like to revisit the core concept of invalidating within the CG-update layer at all. I'm wondering if we would be better served forcing the callers to handle the invalidation beforehand in the cases that they can handle it. An interesting example is when we want to teach the inliner to *update and preserve* analyses. But we can cross that bridge when we get there. With this patch, the new pass manager an build all of the LLVM test suite at -O3 and everything passes. =D I haven't bootstrapped yet and I'm sure there are still plenty of bugs, but this gives a nice baseline so I'm going to increasingly focus on fleshing out the missing functionality, especially the bits that are just turned off right now in order to let us establish this baseline. llvm-svn: 290664
2016-12-28 18:34:50 +08:00
continue;
}
// Now update the call graph.
[PM/LCG] Teach the LazyCallGraph how to replace a function without disturbing the graph or having to update edges. This is motivated by porting argument promotion to the new pass manager. Because of how LLVM IR Function objects work, in order to change their signature a new object needs to be created. This is efficient and straight forward in the IR but previously was very hard to implement in LCG. We could easily replace the function a node in the graph represents. The challenging part is how to handle updating the edges in the graph. LCG previously used an edge to a raw function to represent a node that had not yet been scanned for calls and references. This was the core of its laziness. However, that model causes this kind of update to be very hard: 1) The keys to lookup an edge need to be `Function*`s that would all need to be updated when we update the node. 2) There will be some unknown number of edges that haven't transitioned from `Function*` edges to `Node*` edges. All of this complexity isn't necessary. Instead, we can always build a node around any function, always pointing edges at it and always using it as the key to lookup an edge. To maintain the laziness, we need to sink the *edges* of a node into a secondary object and explicitly model transitioning a node from empty to populated by scanning the function. This design seems much cleaner in a number of ways, but importantly there is now exactly *one* place where the `Function*` has to be updated! Some other cleanups that fall out of this include having something to model the *entry* edges more accurately. Rather than hand rolling parts of the node in the graph itself, we have an explicit `EdgeSequence` object that gives us exactly the functionality needed. We also have a consistent place to define the edge iterators and can use them for both the entry edges and the internal edges of the graph. The API used to model the separation between a node and its edges is intentionally very thin as most clients are expected to deal with nodes that have populated edges. We model this exactly as an optional does with an additional method to populate the edges when that is a reasonable thing for a client to do. This is based on API design suggestions from Richard Smith and David Blaikie, credit goes to them for helping pick how to model this without it being either too explicit or too implicit. The patch is somewhat noisy due to shifting around iterator types and new syntax for walking the edges of a node, but most of the functionality change is in the `Edge`, `EdgeSequence`, and `Node` types. Differential Revision: https://reviews.llvm.org/D29577 llvm-svn: 294653
2017-02-10 07:24:13 +08:00
C = incorporateNewSCCRange(RC->switchInternalEdgeToRef(N, *RefTarget), G, N,
C, AM, UR);
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
}
// We added a ref edge earlier for new call edges, promote those to call edges
// alongside PromotedRefTargets.
for (Node *E : NewCallEdges)
PromotedRefTargets.insert(E);
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// Now promote ref edges into call edges.
[PM/LCG] Teach the LazyCallGraph how to replace a function without disturbing the graph or having to update edges. This is motivated by porting argument promotion to the new pass manager. Because of how LLVM IR Function objects work, in order to change their signature a new object needs to be created. This is efficient and straight forward in the IR but previously was very hard to implement in LCG. We could easily replace the function a node in the graph represents. The challenging part is how to handle updating the edges in the graph. LCG previously used an edge to a raw function to represent a node that had not yet been scanned for calls and references. This was the core of its laziness. However, that model causes this kind of update to be very hard: 1) The keys to lookup an edge need to be `Function*`s that would all need to be updated when we update the node. 2) There will be some unknown number of edges that haven't transitioned from `Function*` edges to `Node*` edges. All of this complexity isn't necessary. Instead, we can always build a node around any function, always pointing edges at it and always using it as the key to lookup an edge. To maintain the laziness, we need to sink the *edges* of a node into a secondary object and explicitly model transitioning a node from empty to populated by scanning the function. This design seems much cleaner in a number of ways, but importantly there is now exactly *one* place where the `Function*` has to be updated! Some other cleanups that fall out of this include having something to model the *entry* edges more accurately. Rather than hand rolling parts of the node in the graph itself, we have an explicit `EdgeSequence` object that gives us exactly the functionality needed. We also have a consistent place to define the edge iterators and can use them for both the entry edges and the internal edges of the graph. The API used to model the separation between a node and its edges is intentionally very thin as most clients are expected to deal with nodes that have populated edges. We model this exactly as an optional does with an additional method to populate the edges when that is a reasonable thing for a client to do. This is based on API design suggestions from Richard Smith and David Blaikie, credit goes to them for helping pick how to model this without it being either too explicit or too implicit. The patch is somewhat noisy due to shifting around iterator types and new syntax for walking the edges of a node, but most of the functionality change is in the `Edge`, `EdgeSequence`, and `Node` types. Differential Revision: https://reviews.llvm.org/D29577 llvm-svn: 294653
2017-02-10 07:24:13 +08:00
for (Node *CallTarget : PromotedRefTargets) {
SCC &TargetC = *G.lookupSCC(*CallTarget);
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
RefSCC &TargetRC = TargetC.getOuterRefSCC();
// The easy case is when the target RefSCC is not this RefSCC. This is
// only supported when the target RefSCC is a child of this RefSCC.
if (&TargetRC != RC) {
assert(RC->isAncestorOf(TargetRC) &&
"Cannot potentially form RefSCC cycles here!");
[PM/LCG] Teach the LazyCallGraph how to replace a function without disturbing the graph or having to update edges. This is motivated by porting argument promotion to the new pass manager. Because of how LLVM IR Function objects work, in order to change their signature a new object needs to be created. This is efficient and straight forward in the IR but previously was very hard to implement in LCG. We could easily replace the function a node in the graph represents. The challenging part is how to handle updating the edges in the graph. LCG previously used an edge to a raw function to represent a node that had not yet been scanned for calls and references. This was the core of its laziness. However, that model causes this kind of update to be very hard: 1) The keys to lookup an edge need to be `Function*`s that would all need to be updated when we update the node. 2) There will be some unknown number of edges that haven't transitioned from `Function*` edges to `Node*` edges. All of this complexity isn't necessary. Instead, we can always build a node around any function, always pointing edges at it and always using it as the key to lookup an edge. To maintain the laziness, we need to sink the *edges* of a node into a secondary object and explicitly model transitioning a node from empty to populated by scanning the function. This design seems much cleaner in a number of ways, but importantly there is now exactly *one* place where the `Function*` has to be updated! Some other cleanups that fall out of this include having something to model the *entry* edges more accurately. Rather than hand rolling parts of the node in the graph itself, we have an explicit `EdgeSequence` object that gives us exactly the functionality needed. We also have a consistent place to define the edge iterators and can use them for both the entry edges and the internal edges of the graph. The API used to model the separation between a node and its edges is intentionally very thin as most clients are expected to deal with nodes that have populated edges. We model this exactly as an optional does with an additional method to populate the edges when that is a reasonable thing for a client to do. This is based on API design suggestions from Richard Smith and David Blaikie, credit goes to them for helping pick how to model this without it being either too explicit or too implicit. The patch is somewhat noisy due to shifting around iterator types and new syntax for walking the edges of a node, but most of the functionality change is in the `Edge`, `EdgeSequence`, and `Node` types. Differential Revision: https://reviews.llvm.org/D29577 llvm-svn: 294653
2017-02-10 07:24:13 +08:00
RC->switchOutgoingEdgeToCall(N, *CallTarget);
LLVM_DEBUG(dbgs() << "Switch outgoing ref edge to a call edge from '" << N
<< "' to '" << *CallTarget << "'\n");
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
continue;
}
LLVM_DEBUG(dbgs() << "Switch an internal ref edge to a call edge from '"
<< N << "' to '" << *CallTarget << "'\n");
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// Otherwise we are switching an internal ref edge to a call edge. This
// may merge away some SCCs, and we add those to the UpdateResult. We also
// need to make sure to update the worklist in the event SCCs have moved
// before the current one in the post-order sequence
bool HasFunctionAnalysisProxy = false;
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
auto InitialSCCIndex = RC->find(*C) - RC->begin();
bool FormedCycle = RC->switchInternalEdgeToCall(
N, *CallTarget, [&](ArrayRef<SCC *> MergedSCCs) {
for (SCC *MergedC : MergedSCCs) {
assert(MergedC != &TargetC && "Cannot merge away the target SCC!");
HasFunctionAnalysisProxy |=
AM.getCachedResult<FunctionAnalysisManagerCGSCCProxy>(
*MergedC) != nullptr;
// Mark that this SCC will no longer be valid.
UR.InvalidatedSCCs.insert(MergedC);
// FIXME: We should really do a 'clear' here to forcibly release
// memory, but we don't have a good way of doing that and
// preserving the function analyses.
auto PA = PreservedAnalyses::allInSet<AllAnalysesOn<Function>>();
PA.preserve<FunctionAnalysisManagerCGSCCProxy>();
AM.invalidate(*MergedC, PA);
}
});
// If we formed a cycle by creating this call, we need to update more data
// structures.
if (FormedCycle) {
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
C = &TargetC;
assert(G.lookupSCC(N) == C && "Failed to update current SCC!");
// If one of the invalidated SCCs had a cached proxy to a function
// analysis manager, we need to create a proxy in the new current SCC as
// the invalidated SCCs had their functions moved.
if (HasFunctionAnalysisProxy)
AM.getResult<FunctionAnalysisManagerCGSCCProxy>(*C, G).updateFAM(FAM);
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// Any analyses cached for this SCC are no longer precise as the shape
// has changed by introducing this cycle. However, we have taken care to
// update the proxies so it remains valide.
auto PA = PreservedAnalyses::allInSet<AllAnalysesOn<Function>>();
PA.preserve<FunctionAnalysisManagerCGSCCProxy>();
AM.invalidate(*C, PA);
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
}
auto NewSCCIndex = RC->find(*C) - RC->begin();
// If we have actually moved an SCC to be topologically "below" the current
// one due to merging, we will need to revisit the current SCC after
// visiting those moved SCCs.
//
// It is critical that we *do not* revisit the current SCC unless we
// actually move SCCs in the process of merging because otherwise we may
// form a cycle where an SCC is split apart, merged, split, merged and so
// on infinitely.
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
if (InitialSCCIndex < NewSCCIndex) {
// Put our current SCC back onto the worklist as we'll visit other SCCs
// that are now definitively ordered prior to the current one in the
// post-order sequence, and may end up observing more precise context to
// optimize the current SCC.
UR.CWorklist.insert(C);
LLVM_DEBUG(dbgs() << "Enqueuing the existing SCC in the worklist: " << *C
<< "\n");
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
// Enqueue in reverse order as we pop off the back of the worklist.
for (SCC &MovedC : llvm::reverse(make_range(RC->begin() + InitialSCCIndex,
RC->begin() + NewSCCIndex))) {
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
UR.CWorklist.insert(&MovedC);
LLVM_DEBUG(dbgs() << "Enqueuing a newly earlier in post-order SCC: "
<< MovedC << "\n");
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
}
}
}
assert(!UR.InvalidatedSCCs.count(C) && "Invalidated the current SCC!");
assert(!UR.InvalidatedRefSCCs.count(RC) && "Invalidated the current RefSCC!");
assert(&C->getOuterRefSCC() == RC && "Current SCC not in current RefSCC!");
// Record the current RefSCC and SCC for higher layers of the CGSCC pass
// manager now that all the updates have been applied.
if (RC != &InitialRC)
UR.UpdatedRC = RC;
if (C != &InitialC)
UR.UpdatedC = C;
return *C;
}
LazyCallGraph::SCC &llvm::updateCGAndAnalysisManagerForFunctionPass(
LazyCallGraph &G, LazyCallGraph::SCC &InitialC, LazyCallGraph::Node &N,
CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR,
FunctionAnalysisManager &FAM) {
return updateCGAndAnalysisManagerForPass(G, InitialC, N, AM, UR, FAM,
/* FunctionPass */ true);
}
LazyCallGraph::SCC &llvm::updateCGAndAnalysisManagerForCGSCCPass(
LazyCallGraph &G, LazyCallGraph::SCC &InitialC, LazyCallGraph::Node &N,
CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR,
FunctionAnalysisManager &FAM) {
return updateCGAndAnalysisManagerForPass(G, InitialC, N, AM, UR, FAM,
/* FunctionPass */ false);
}