1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-22 04:22:57 +02:00
llvm-mirror/include/llvm/Analysis/CGSCCPassManager.h

841 lines
35 KiB
C
Raw Normal View History

//===- CGSCCPassManager.h - Call graph pass management ----------*- C++ -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
/// \file
///
/// This header provides classes for managing passes over SCCs of the call
/// graph. These passes form an important component of LLVM's interprocedural
/// optimizations. Because they operate on the SCCs of the call graph, and they
[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 11:37:14 +02:00
/// traverse the graph in post-order, they can effectively do pair-wise
/// interprocedural optimizations for all call edges in the program while
/// incrementally refining it and improving the context of these pair-wise
/// optimizations. At each call site edge, the callee has already been
/// optimized as much as is possible. This in turn allows very accurate
/// analysis of it for IPO.
///
/// A secondary more general goal is to be able to isolate optimization on
/// unrelated parts of the IR module. This is useful to ensure our
/// optimizations are principled and don't miss oportunities where refinement
/// of one part of the module influence transformations in another part of the
/// module. But this is also useful if we want to parallelize the optimizations
/// across common large module graph shapes which tend to be very wide and have
/// large regions of unrelated cliques.
///
/// To satisfy these goals, we use the LazyCallGraph which provides two graphs
/// nested inside each other (and built lazily from the bottom-up): the call
/// graph proper, and a reference graph. The reference graph is super set of
/// the call graph and is a conservative approximation of what could through
/// scalar or CGSCC transforms *become* the call graph. Using this allows us to
/// ensure we optimize functions prior to them being introduced into the call
/// graph by devirtualization or other technique, and thus ensures that
/// subsequent pair-wise interprocedural optimizations observe the optimized
/// form of these functions. The (potentially transitive) reference
/// reachability used by the reference graph is a conservative approximation
/// that still allows us to have independent regions of the graph.
///
/// FIXME: There is one major drawback of the reference graph: in its naive
/// form it is quadratic because it contains a distinct edge for each
/// (potentially indirect) reference, even if are all through some common
/// global table of function pointers. This can be fixed in a number of ways
/// that essentially preserve enough of the normalization. While it isn't
/// expected to completely preclude the usability of this, it will need to be
/// addressed.
///
///
/// All of these issues are made substantially more complex in the face of
/// mutations to the call graph while optimization passes are being run. When
/// mutations to the call graph occur we want to achieve two different things:
///
/// - We need to update the call graph in-flight and invalidate analyses
/// cached on entities in the graph. Because of the cache-based analysis
/// design of the pass manager, it is essential to have stable identities for
/// the elements of the IR that passes traverse, and to invalidate any
/// analyses cached on these elements as the mutations take place.
///
/// - We want to preserve the incremental and post-order traversal of the
/// graph even as it is refined and mutated. This means we want optimization
/// to observe the most refined form of the call graph and to do so in
/// post-order.
///
/// To address this, the CGSCC manager uses both worklists that can be expanded
/// by passes which transform the IR, and provides invalidation tests to skip
/// entries that become dead. This extra data is provided to every SCC pass so
/// that it can carefully update the manager's traversal as the call graph
/// mutates.
///
/// We also provide support for running function passes within the CGSCC walk,
/// and there we provide automatic update of the call graph including of the
/// pass manager to reflect call graph changes that fall out naturally as part
/// of scalar transformations.
///
/// The patterns used to ensure the goals of post-order visitation of the fully
/// refined graph:
///
/// 1) Sink toward the "bottom" as the graph is refined. This means that any
/// iteration continues in some valid post-order sequence after the mutation
/// has altered the structure.
///
/// 2) Enqueue in post-order, including the current entity. If the current
/// entity's shape changes, it and everything after it in post-order needs
/// to be visited to observe that shape.
///
//===----------------------------------------------------------------------===//
#ifndef LLVM_ANALYSIS_CGSCCPASSMANAGER_H
#define LLVM_ANALYSIS_CGSCCPASSMANAGER_H
#include "llvm/ADT/DenseSet.h"
[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 11:37:14 +02:00
#include "llvm/ADT/PriorityWorklist.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/Analysis/LazyCallGraph.h"
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
#include "llvm/IR/CallSite.h"
#include "llvm/IR/Function.h"
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/PassManager.h"
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
#include "llvm/IR/ValueHandle.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include <algorithm>
#include <cassert>
#include <utility>
namespace llvm {
struct CGSCCUpdateResult;
class Module;
// Allow debug logging in this inline function.
#define DEBUG_TYPE "cgscc"
[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
2016-11-28 23:04:31 +01:00
/// Extern template declaration for the analysis set for this IR unit.
extern 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 11:37:14 +02:00
extern template class AnalysisManager<LazyCallGraph::SCC, LazyCallGraph &>;
/// The CGSCC analysis manager.
///
/// See the documentation for the AnalysisManager template for detail
/// documentation. This type serves as a convenient way to refer to this
/// construct in the adaptors and proxies used to integrate this into the larger
/// pass manager infrastructure.
using CGSCCAnalysisManager =
AnalysisManager<LazyCallGraph::SCC, LazyCallGraph &>;
[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 11:37:14 +02:00
// Explicit specialization and instantiation declarations for the pass manager.
// See the comments on the definition of the specialization for details on how
// it differs from the primary template.
template <>
PreservedAnalyses
PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager, LazyCallGraph &,
CGSCCUpdateResult &>::run(LazyCallGraph::SCC &InitialC,
CGSCCAnalysisManager &AM,
LazyCallGraph &G, CGSCCUpdateResult &UR);
extern template class PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager,
LazyCallGraph &, CGSCCUpdateResult &>;
/// The CGSCC pass manager.
[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 11:37:14 +02:00
///
/// See the documentation for the PassManager template for details. It runs
/// a sequence of SCC passes over each SCC that the manager is run over. This
/// type serves as a convenient way to refer to this construct.
using CGSCCPassManager =
PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager, LazyCallGraph &,
CGSCCUpdateResult &>;
[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 11:37:14 +02:00
/// An explicit specialization of the require analysis template pass.
template <typename AnalysisT>
struct RequireAnalysisPass<AnalysisT, LazyCallGraph::SCC, CGSCCAnalysisManager,
LazyCallGraph &, CGSCCUpdateResult &>
: PassInfoMixin<RequireAnalysisPass<AnalysisT, LazyCallGraph::SCC,
CGSCCAnalysisManager, LazyCallGraph &,
CGSCCUpdateResult &>> {
PreservedAnalyses run(LazyCallGraph::SCC &C, CGSCCAnalysisManager &AM,
LazyCallGraph &CG, CGSCCUpdateResult &) {
(void)AM.template getResult<AnalysisT>(C, CG);
return PreservedAnalyses::all();
}
};
/// A proxy from a \c CGSCCAnalysisManager to a \c Module.
using CGSCCAnalysisManagerModuleProxy =
InnerAnalysisManagerProxy<CGSCCAnalysisManager, 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 07:34:44 +01:00
/// We need a specialized result for the \c CGSCCAnalysisManagerModuleProxy so
/// it can have access to the call graph in order to walk all the SCCs when
/// invalidating things.
template <> class CGSCCAnalysisManagerModuleProxy::Result {
public:
explicit Result(CGSCCAnalysisManager &InnerAM, LazyCallGraph &G)
: InnerAM(&InnerAM), G(&G) {}
/// Accessor for the analysis manager.
[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 07:34:44 +01:00
CGSCCAnalysisManager &getManager() { return *InnerAM; }
/// Handler for invalidation of the 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 07:34:44 +01:00
///
/// If the proxy analysis itself is preserved, then we assume that the set of
/// SCCs in the Module hasn't changed. Thus any pointers to SCCs in the
/// CGSCCAnalysisManager are still valid, and we don't need to call \c clear
/// on the CGSCCAnalysisManager.
///
/// Regardless of whether this analysis is marked as preserved, all of the
/// analyses in the \c CGSCCAnalysisManager are potentially invalidated based
/// on the set of preserved analyses.
bool invalidate(Module &M, const PreservedAnalyses &PA,
ModuleAnalysisManager::Invalidator &Inv);
private:
CGSCCAnalysisManager *InnerAM;
LazyCallGraph *G;
};
/// Provide a specialized run method for the \c CGSCCAnalysisManagerModuleProxy
/// so it can pass the lazy call graph to the result.
template <>
CGSCCAnalysisManagerModuleProxy::Result
CGSCCAnalysisManagerModuleProxy::run(Module &M, ModuleAnalysisManager &AM);
// Ensure the \c CGSCCAnalysisManagerModuleProxy is provided as an extern
// template.
extern template class InnerAnalysisManagerProxy<CGSCCAnalysisManager, Module>;
extern template class OuterAnalysisManagerProxy<
ModuleAnalysisManager, LazyCallGraph::SCC, LazyCallGraph &>;
/// A proxy from a \c ModuleAnalysisManager to an \c SCC.
using ModuleAnalysisManagerCGSCCProxy =
OuterAnalysisManagerProxy<ModuleAnalysisManager, LazyCallGraph::SCC,
LazyCallGraph &>;
[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 11:37:14 +02:00
/// Support structure for SCC passes to communicate updates the call graph back
/// to the CGSCC pass manager infrsatructure.
///
/// The CGSCC pass manager runs SCC passes which are allowed to update the call
/// graph and SCC structures. This means the structure the pass manager works
/// on is mutating underneath it. In order to support that, there needs to be
/// careful communication about the precise nature and ramifications of these
/// updates to the pass management infrastructure.
///
/// All SCC passes will have to accept a reference to the management layer's
/// update result struct and use it to reflect the results of any CG updates
/// performed.
///
/// Passes which do not change the call graph structure in any way can just
/// ignore this argument to their run method.
struct CGSCCUpdateResult {
/// Worklist of the RefSCCs queued for processing.
///
/// When a pass refines the graph and creates new RefSCCs or causes them to
/// have a different shape or set of component SCCs it should add the RefSCCs
/// to this worklist so that we visit them in the refined form.
///
/// This worklist is in reverse post-order, as we pop off the back in order
/// to observe RefSCCs in post-order. When adding RefSCCs, clients should add
/// them in reverse post-order.
SmallPriorityWorklist<LazyCallGraph::RefSCC *, 1> &RCWorklist;
/// Worklist of the SCCs queued for processing.
///
/// When a pass refines the graph and creates new SCCs or causes them to have
/// a different shape or set of component functions it should add the SCCs to
/// this worklist so that we visit them in the refined form.
///
/// Note that if the SCCs are part of a RefSCC that is added to the \c
/// RCWorklist, they don't need to be added here as visiting the RefSCC will
/// be sufficient to re-visit the SCCs within it.
///
/// This worklist is in reverse post-order, as we pop off the back in order
/// to observe SCCs in post-order. When adding SCCs, clients should add them
/// in reverse post-order.
SmallPriorityWorklist<LazyCallGraph::SCC *, 1> &CWorklist;
/// The set of invalidated RefSCCs which should be skipped if they are found
/// in \c RCWorklist.
///
/// This is used to quickly prune out RefSCCs when they get deleted and
/// happen to already be on the worklist. We use this primarily to avoid
/// scanning the list and removing entries from it.
SmallPtrSetImpl<LazyCallGraph::RefSCC *> &InvalidatedRefSCCs;
/// The set of invalidated SCCs which should be skipped if they are found
/// in \c CWorklist.
///
/// This is used to quickly prune out SCCs when they get deleted and happen
/// to already be on the worklist. We use this primarily to avoid scanning
/// the list and removing entries from it.
SmallPtrSetImpl<LazyCallGraph::SCC *> &InvalidatedSCCs;
/// If non-null, the updated current \c RefSCC being processed.
///
/// This is set when a graph refinement takes place an the "current" point in
/// the graph moves "down" or earlier in the post-order walk. This will often
/// cause the "current" RefSCC to be a newly created RefSCC object and the
/// old one to be added to the above worklist. When that happens, this
/// pointer is non-null and can be used to continue processing the "top" of
/// the post-order walk.
LazyCallGraph::RefSCC *UpdatedRC;
/// If non-null, the updated current \c SCC being processed.
///
/// This is set when a graph refinement takes place an the "current" point in
/// the graph moves "down" or earlier in the post-order walk. This will often
/// cause the "current" SCC to be a newly created SCC object and the old one
/// to be added to the above worklist. When that happens, this pointer is
/// non-null and can be used to continue processing the "top" of the
/// post-order walk.
LazyCallGraph::SCC *UpdatedC;
[PM] Fix a bug where through CGSCC iteration we can get infinite-inlining across multiple runs of the inliner by keeping a tiny history of internal-to-SCC inlining decisions. This is still a bit gross, but I don't yet have any fundamentally better ideas and numerous people are blocked on this to use new PM and ThinLTO together. The core of the idea is to detect when we are about to do an inline that has a chance of re-splitting an SCC which we have split before with a similar inlining step. That is a critical component in the inlining forming a cycle and so far detects all of the various cyclic patterns I can come up with as well as the original real-world test case (which comes from a ThinLTO build of libunwind). I've added some tests that I think really demonstrate what is going on here. They are essentially state machines that march the inliner through various steps of a cycle and check that we stop when the cycle is closed and that we actually did do inlining to form that cycle. A lot of thanks go to Eric Christopher and Sanjoy Das for the help understanding this issue and improving the test cases. The biggest "yuck" here is the layering issue -- the CGSCC pass manager is providing somewhat magical state to the inliner for it to use to make itself converge. This isn't great, but I don't honestly have a lot of better ideas yet and at least seems nicely isolated. I have tested this patch, and it doesn't block *any* inlining on the entire LLVM test suite and SPEC, so it seems sufficiently narrowly targeted to the issue at hand. We have come up with hypothetical issues that this patch doesn't cover, but so far none of them are practical and we don't have a viable solution yet that covers the hypothetical stuff, so proceeding here in the interim. Definitely an area that we will be back and revisiting in the future. Differential Revision: https://reviews.llvm.org/D36188 llvm-svn: 309784
2017-08-02 04:09:22 +02:00
/// A hacky area where the inliner can retain history about inlining
/// decisions that mutated the call graph's SCC structure in order to avoid
/// infinite inlining. See the comments in the inliner's CG update logic.
///
/// FIXME: Keeping this here seems like a big layering issue, we should look
/// for a better technique.
SmallDenseSet<std::pair<LazyCallGraph::Node *, LazyCallGraph::SCC *>, 4>
&InlinedInternalEdges;
[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 11:37:14 +02:00
};
/// The core module pass which does a post-order walk of the SCCs and
/// runs a CGSCC pass over each one.
///
/// Designed to allow composition of a CGSCCPass(Manager) and
/// a ModulePassManager. Note that this pass must be run with a module analysis
/// manager as it uses the LazyCallGraph analysis. It will also run the
/// \c CGSCCAnalysisManagerModuleProxy analysis prior to running the CGSCC
/// pass over the module to enable a \c FunctionAnalysisManager to be used
/// within this run safely.
template <typename CGSCCPassT>
class ModuleToPostOrderCGSCCPassAdaptor
: public PassInfoMixin<ModuleToPostOrderCGSCCPassAdaptor<CGSCCPassT>> {
public:
explicit ModuleToPostOrderCGSCCPassAdaptor(CGSCCPassT Pass)
: Pass(std::move(Pass)) {}
// We have to explicitly define all the special member functions because MSVC
// refuses to generate them.
ModuleToPostOrderCGSCCPassAdaptor(
const ModuleToPostOrderCGSCCPassAdaptor &Arg)
: Pass(Arg.Pass) {}
ModuleToPostOrderCGSCCPassAdaptor(ModuleToPostOrderCGSCCPassAdaptor &&Arg)
: Pass(std::move(Arg.Pass)) {}
friend void swap(ModuleToPostOrderCGSCCPassAdaptor &LHS,
ModuleToPostOrderCGSCCPassAdaptor &RHS) {
std::swap(LHS.Pass, RHS.Pass);
}
ModuleToPostOrderCGSCCPassAdaptor &
operator=(ModuleToPostOrderCGSCCPassAdaptor RHS) {
swap(*this, RHS);
return *this;
}
/// Runs the CGSCC pass across every SCC in the module.
PreservedAnalyses 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);
[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 11:37:14 +02:00
// 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;
[PM] Fix a bug where through CGSCC iteration we can get infinite-inlining across multiple runs of the inliner by keeping a tiny history of internal-to-SCC inlining decisions. This is still a bit gross, but I don't yet have any fundamentally better ideas and numerous people are blocked on this to use new PM and ThinLTO together. The core of the idea is to detect when we are about to do an inline that has a chance of re-splitting an SCC which we have split before with a similar inlining step. That is a critical component in the inlining forming a cycle and so far detects all of the various cyclic patterns I can come up with as well as the original real-world test case (which comes from a ThinLTO build of libunwind). I've added some tests that I think really demonstrate what is going on here. They are essentially state machines that march the inliner through various steps of a cycle and check that we stop when the cycle is closed and that we actually did do inlining to form that cycle. A lot of thanks go to Eric Christopher and Sanjoy Das for the help understanding this issue and improving the test cases. The biggest "yuck" here is the layering issue -- the CGSCC pass manager is providing somewhat magical state to the inliner for it to use to make itself converge. This isn't great, but I don't honestly have a lot of better ideas yet and at least seems nicely isolated. I have tested this patch, and it doesn't block *any* inlining on the entire LLVM test suite and SPEC, so it seems sufficiently narrowly targeted to the issue at hand. We have come up with hypothetical issues that this patch doesn't cover, but so far none of them are practical and we don't have a viable solution yet that covers the hypothetical stuff, so proceeding here in the interim. Definitely an area that we will be back and revisiting in the future. Differential Revision: https://reviews.llvm.org/D36188 llvm-svn: 309784
2017-08-02 04:09:22 +02:00
SmallDenseSet<std::pair<LazyCallGraph::Node *, LazyCallGraph::SCC *>, 4>
InlinedInternalEdges;
CGSCCUpdateResult UR = {RCWorklist, CWorklist, InvalidRefSCCSet,
InvalidSCCSet, nullptr, nullptr,
InlinedInternalEdges};
[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 11:37:14 +02:00
PreservedAnalyses PA = PreservedAnalyses::all();
[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
2017-02-06 20:38:06 +01:00
CG.buildRefSCCs();
[PM] Provide an initial, minimal port of the inliner to the new pass manager. This doesn't implement *every* feature of the existing inliner, but tries to implement the most important ones for building a functional optimization pipeline and beginning to sort out bugs, regressions, and other problems. Notable, but intentional omissions: - No alloca merging support. Why? Because it isn't clear we want to do this at all. Active discussion and investigation is going on to remove it, so for simplicity I omitted it. - No support for trying to iterate on "internally" devirtualized calls. Why? Because it adds what I suspect is inappropriate coupling for little or no benefit. We will have an outer iteration system that tracks devirtualization including that from function passes and iterates already. We should improve that rather than approximate it here. - Optimization remarks. Why? Purely to make the patch smaller, no other reason at all. The last one I'll probably work on almost immediately. But I wanted to skip it in the initial patch to try to focus the change as much as possible as there is already a lot of code moving around and both of these *could* be skipped without really disrupting the core logic. A summary of the different things happening here: 1) Adding the usual new PM class and rigging. 2) Fixing minor underlying assumptions in the inline cost analysis or inline logic that don't generally hold in the new PM world. 3) Adding the core pass logic which is in essence a loop over the calls in the nodes in the call graph. This is a bit duplicated from the old inliner, but only a handful of lines could realistically be shared. (I tried at first, and it really didn't help anything.) All told, this is only about 100 lines of code, and most of that is the mechanics of wiring up analyses from the new PM world. 4) Updating the LazyCallGraph (in the new PM) based on the *newly inlined* calls and references. This is very minimal because we cannot form cycles. 5) When inlining removes the last use of a function, eagerly nuking the body of the function so that any "one use remaining" inline cost heuristics are immediately refined, and queuing these functions to be completely deleted once inlining is complete and the call graph updated to reflect that they have become dead. 6) After all the inlining for a particular function, updating the LazyCallGraph and the CGSCC pass manager to reflect the function-local simplifications that are done immediately and internally by the inline utilties. These are the exact same fundamental set of CG updates done by arbitrary function passes. 7) Adding a bunch of test cases to specifically target CGSCC and other subtle aspects in the new PM world. Many thanks to the careful review from Easwaran and Sanjoy and others! Differential Revision: https://reviews.llvm.org/D24226 llvm-svn: 290161
2016-12-20 04:15:32 +01:00
for (auto RCI = CG.postorder_ref_scc_begin(),
RCE = CG.postorder_ref_scc_end();
RCI != RCE;) {
[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 11:37:14 +02:00
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.
[PM] Provide an initial, minimal port of the inliner to the new pass manager. This doesn't implement *every* feature of the existing inliner, but tries to implement the most important ones for building a functional optimization pipeline and beginning to sort out bugs, regressions, and other problems. Notable, but intentional omissions: - No alloca merging support. Why? Because it isn't clear we want to do this at all. Active discussion and investigation is going on to remove it, so for simplicity I omitted it. - No support for trying to iterate on "internally" devirtualized calls. Why? Because it adds what I suspect is inappropriate coupling for little or no benefit. We will have an outer iteration system that tracks devirtualization including that from function passes and iterates already. We should improve that rather than approximate it here. - Optimization remarks. Why? Purely to make the patch smaller, no other reason at all. The last one I'll probably work on almost immediately. But I wanted to skip it in the initial patch to try to focus the change as much as possible as there is already a lot of code moving around and both of these *could* be skipped without really disrupting the core logic. A summary of the different things happening here: 1) Adding the usual new PM class and rigging. 2) Fixing minor underlying assumptions in the inline cost analysis or inline logic that don't generally hold in the new PM world. 3) Adding the core pass logic which is in essence a loop over the calls in the nodes in the call graph. This is a bit duplicated from the old inliner, but only a handful of lines could realistically be shared. (I tried at first, and it really didn't help anything.) All told, this is only about 100 lines of code, and most of that is the mechanics of wiring up analyses from the new PM world. 4) Updating the LazyCallGraph (in the new PM) based on the *newly inlined* calls and references. This is very minimal because we cannot form cycles. 5) When inlining removes the last use of a function, eagerly nuking the body of the function so that any "one use remaining" inline cost heuristics are immediately refined, and queuing these functions to be completely deleted once inlining is complete and the call graph updated to reflect that they have become dead. 6) After all the inlining for a particular function, updating the LazyCallGraph and the CGSCC pass manager to reflect the function-local simplifications that are done immediately and internally by the inline utilties. These are the exact same fundamental set of CG updates done by arbitrary function passes. 7) Adding a bunch of test cases to specifically target CGSCC and other subtle aspects in the new PM world. Many thanks to the careful review from Easwaran and Sanjoy and others! Differential Revision: https://reviews.llvm.org/D24226 llvm-svn: 290161
2016-12-20 04:15:32 +01:00
//
// We also eagerly increment the iterator to the next position because
// the CGSCC passes below may delete the current RefSCC.
RCWorklist.insert(&*RCI++);
[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 11:37:14 +02:00
do {
LazyCallGraph::RefSCC *RC = RCWorklist.pop_back_val();
if (InvalidRefSCCSet.count(RC)) {
DEBUG(dbgs() << "Skipping an invalid RefSCC...\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 11:37:14 +02:00
continue;
}
[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 11:37:14 +02:00
assert(CWorklist.empty() &&
"Should always start with an empty SCC worklist");
DEBUG(dbgs() << "Running an SCC pass across the RefSCC: " << *RC
<< "\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 11:37:14 +02:00
// Push the initial SCCs in reverse post-order as we'll pop off the
[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 11:37:14 +02:00
// back and so see this in post-order.
for (LazyCallGraph::SCC &C : llvm::reverse(*RC))
[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 11:37:14 +02:00
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)) {
DEBUG(dbgs() << "Skipping an invalid SCC...\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 11:37:14 +02:00
continue;
}
if (&C->getOuterRefSCC() != RC) {
DEBUG(dbgs() << "Skipping an SCC that is now part of some other "
"RefSCC...\n");
continue;
}
[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 11:37:14 +02:00
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!");
UR.UpdatedRC = nullptr;
UR.UpdatedC = nullptr;
PreservedAnalyses PassPA = Pass.run(*C, CGAM, CG, UR);
// Update the SCC and RefSCC if necessary.
C = UR.UpdatedC ? UR.UpdatedC : C;
RC = UR.UpdatedRC ? UR.UpdatedRC : RC;
// 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)) {
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!");
[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 11:37:14 +02:00
// 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);
[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 11:37:14 +02:00
// Then intersect the preserved set so that invalidation of module
// analyses will eventually occur when the module pass completes.
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)
DEBUG(dbgs() << "Re-running SCC passes after a refinement of the "
"current SCC: "
<< *UR.UpdatedC << "\n");
[PM] Provide an initial, minimal port of the inliner to the new pass manager. This doesn't implement *every* feature of the existing inliner, but tries to implement the most important ones for building a functional optimization pipeline and beginning to sort out bugs, regressions, and other problems. Notable, but intentional omissions: - No alloca merging support. Why? Because it isn't clear we want to do this at all. Active discussion and investigation is going on to remove it, so for simplicity I omitted it. - No support for trying to iterate on "internally" devirtualized calls. Why? Because it adds what I suspect is inappropriate coupling for little or no benefit. We will have an outer iteration system that tracks devirtualization including that from function passes and iterates already. We should improve that rather than approximate it here. - Optimization remarks. Why? Purely to make the patch smaller, no other reason at all. The last one I'll probably work on almost immediately. But I wanted to skip it in the initial patch to try to focus the change as much as possible as there is already a lot of code moving around and both of these *could* be skipped without really disrupting the core logic. A summary of the different things happening here: 1) Adding the usual new PM class and rigging. 2) Fixing minor underlying assumptions in the inline cost analysis or inline logic that don't generally hold in the new PM world. 3) Adding the core pass logic which is in essence a loop over the calls in the nodes in the call graph. This is a bit duplicated from the old inliner, but only a handful of lines could realistically be shared. (I tried at first, and it really didn't help anything.) All told, this is only about 100 lines of code, and most of that is the mechanics of wiring up analyses from the new PM world. 4) Updating the LazyCallGraph (in the new PM) based on the *newly inlined* calls and references. This is very minimal because we cannot form cycles. 5) When inlining removes the last use of a function, eagerly nuking the body of the function so that any "one use remaining" inline cost heuristics are immediately refined, and queuing these functions to be completely deleted once inlining is complete and the call graph updated to reflect that they have become dead. 6) After all the inlining for a particular function, updating the LazyCallGraph and the CGSCC pass manager to reflect the function-local simplifications that are done immediately and internally by the inline utilties. These are the exact same fundamental set of CG updates done by arbitrary function passes. 7) Adding a bunch of test cases to specifically target CGSCC and other subtle aspects in the new PM world. Many thanks to the careful review from Easwaran and Sanjoy and others! Differential Revision: https://reviews.llvm.org/D24226 llvm-svn: 290161
2016-12-20 04:15:32 +01:00
// 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.
[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 11:37:14 +02:00
} while (UR.UpdatedC);
} while (!CWorklist.empty());
[PM] Fix a bug where through CGSCC iteration we can get infinite-inlining across multiple runs of the inliner by keeping a tiny history of internal-to-SCC inlining decisions. This is still a bit gross, but I don't yet have any fundamentally better ideas and numerous people are blocked on this to use new PM and ThinLTO together. The core of the idea is to detect when we are about to do an inline that has a chance of re-splitting an SCC which we have split before with a similar inlining step. That is a critical component in the inlining forming a cycle and so far detects all of the various cyclic patterns I can come up with as well as the original real-world test case (which comes from a ThinLTO build of libunwind). I've added some tests that I think really demonstrate what is going on here. They are essentially state machines that march the inliner through various steps of a cycle and check that we stop when the cycle is closed and that we actually did do inlining to form that cycle. A lot of thanks go to Eric Christopher and Sanjoy Das for the help understanding this issue and improving the test cases. The biggest "yuck" here is the layering issue -- the CGSCC pass manager is providing somewhat magical state to the inliner for it to use to make itself converge. This isn't great, but I don't honestly have a lot of better ideas yet and at least seems nicely isolated. I have tested this patch, and it doesn't block *any* inlining on the entire LLVM test suite and SPEC, so it seems sufficiently narrowly targeted to the issue at hand. We have come up with hypothetical issues that this patch doesn't cover, but so far none of them are practical and we don't have a viable solution yet that covers the hypothetical stuff, so proceeding here in the interim. Definitely an area that we will be back and revisiting in the future. Differential Revision: https://reviews.llvm.org/D36188 llvm-svn: 309784
2017-08-02 04:09:22 +02:00
// 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();
[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 11:37:14 +02:00
} while (!RCWorklist.empty());
}
[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 07:34:44 +01:00
// By definition we preserve the call garph, all SCC analyses, and the
// analysis proxies by handling them above and in any nested pass managers.
[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 09:40:39 +01:00
PA.preserveSet<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 07:34:44 +01:00
PA.preserve<LazyCallGraphAnalysis>();
PA.preserve<CGSCCAnalysisManagerModuleProxy>();
[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 07:34:44 +01:00
PA.preserve<FunctionAnalysisManagerModuleProxy>();
return PA;
}
private:
CGSCCPassT Pass;
};
/// A function to deduce a function pass type and wrap it in the
/// templated adaptor.
template <typename CGSCCPassT>
ModuleToPostOrderCGSCCPassAdaptor<CGSCCPassT>
createModuleToPostOrderCGSCCPassAdaptor(CGSCCPassT Pass) {
return ModuleToPostOrderCGSCCPassAdaptor<CGSCCPassT>(std::move(Pass));
}
/// A proxy from a \c FunctionAnalysisManager to an \c 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 07:34:44 +01:00
///
/// When a module pass runs and triggers invalidation, both the CGSCC and
/// Function analysis manager proxies on the module get an invalidation event.
/// We don't want to fully duplicate responsibility for most of the
/// invalidation logic. Instead, this layer is only responsible for SCC-local
/// invalidation events. We work with the module's FunctionAnalysisManager to
/// invalidate function analyses.
class FunctionAnalysisManagerCGSCCProxy
: public AnalysisInfoMixin<FunctionAnalysisManagerCGSCCProxy> {
public:
class Result {
public:
explicit Result(FunctionAnalysisManager &FAM) : FAM(&FAM) {}
/// Accessor for the analysis manager.
[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 07:34:44 +01:00
FunctionAnalysisManager &getManager() { return *FAM; }
bool invalidate(LazyCallGraph::SCC &C, const PreservedAnalyses &PA,
CGSCCAnalysisManager::Invalidator &Inv);
private:
FunctionAnalysisManager *FAM;
};
/// Computes the \c FunctionAnalysisManager and stores it in the result proxy.
Result run(LazyCallGraph::SCC &C, CGSCCAnalysisManager &AM, LazyCallGraph &);
private:
friend AnalysisInfoMixin<FunctionAnalysisManagerCGSCCProxy>;
[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 07:34:44 +01:00
static AnalysisKey Key;
};
extern template class OuterAnalysisManagerProxy<CGSCCAnalysisManager, Function>;
/// A proxy from a \c CGSCCAnalysisManager to a \c Function.
using CGSCCAnalysisManagerFunctionProxy =
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 11:37:14 +02:00
/// Helper to update the call graph after running a function pass.
///
/// Function passes can only mutate the call graph in specific ways. This
/// routine provides a helper that updates the call graph in those ways
/// including returning whether any changes were made and populating a CG
/// update result struct for the overall CGSCC walk.
LazyCallGraph::SCC &updateCGAndAnalysisManagerForFunctionPass(
LazyCallGraph &G, LazyCallGraph::SCC &C, LazyCallGraph::Node &N,
CGSCCAnalysisManager &AM, CGSCCUpdateResult &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 11:37:14 +02:00
/// Adaptor that maps from a SCC to its functions.
///
/// Designed to allow composition of a FunctionPass(Manager) and
/// a CGSCCPassManager. Note that if this pass is constructed with a pointer
/// to a \c CGSCCAnalysisManager it will run the
/// \c FunctionAnalysisManagerCGSCCProxy analysis prior to running the function
/// pass over the SCC to enable a \c FunctionAnalysisManager to be used
/// within this run safely.
template <typename FunctionPassT>
class CGSCCToFunctionPassAdaptor
: public PassInfoMixin<CGSCCToFunctionPassAdaptor<FunctionPassT>> {
public:
explicit CGSCCToFunctionPassAdaptor(FunctionPassT Pass)
: Pass(std::move(Pass)) {}
// We have to explicitly define all the special member functions because MSVC
// refuses to generate them.
CGSCCToFunctionPassAdaptor(const CGSCCToFunctionPassAdaptor &Arg)
: Pass(Arg.Pass) {}
CGSCCToFunctionPassAdaptor(CGSCCToFunctionPassAdaptor &&Arg)
: Pass(std::move(Arg.Pass)) {}
friend void swap(CGSCCToFunctionPassAdaptor &LHS,
CGSCCToFunctionPassAdaptor &RHS) {
std::swap(LHS.Pass, RHS.Pass);
}
CGSCCToFunctionPassAdaptor &operator=(CGSCCToFunctionPassAdaptor RHS) {
swap(*this, RHS);
return *this;
}
/// Runs the function pass across every function in the module.
[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 11:37:14 +02:00
PreservedAnalyses run(LazyCallGraph::SCC &C, CGSCCAnalysisManager &AM,
LazyCallGraph &CG, CGSCCUpdateResult &UR) {
// Setup the function analysis manager from its proxy.
FunctionAnalysisManager &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 11:37:14 +02:00
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;
DEBUG(dbgs() << "Running function passes across an SCC: " << C << "\n");
PreservedAnalyses PA = PreservedAnalyses::all();
[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 11:37:14 +02:00
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;
PreservedAnalyses PassPA = Pass.run(N->getFunction(), FAM);
// 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(N->getFunction(), 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));
[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 11:37:14 +02:00
// 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);
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.
[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 09:40:39 +01:00
PA.preserveSet<AllAnalysesOn<Function>>();
PA.preserve<FunctionAnalysisManagerCGSCCProxy>();
[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 11:37:14 +02:00
// We've also ensured that we updated the call graph along the way.
PA.preserve<LazyCallGraphAnalysis>();
return PA;
}
private:
FunctionPassT Pass;
};
/// A function to deduce a function pass type and wrap it in the
/// templated adaptor.
template <typename FunctionPassT>
CGSCCToFunctionPassAdaptor<FunctionPassT>
createCGSCCToFunctionPassAdaptor(FunctionPassT Pass) {
return CGSCCToFunctionPassAdaptor<FunctionPassT>(std::move(Pass));
}
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
/// A helper that repeats an SCC pass each time an indirect call is refined to
/// a direct call by that pass.
///
/// While the CGSCC pass manager works to re-visit SCCs and RefSCCs as they
/// change shape, we may also want to repeat an SCC pass if it simply refines
/// an indirect call to a direct call, even if doing so does not alter the
/// shape of the graph. Note that this only pertains to direct calls to
/// functions where IPO across the SCC may be able to compute more precise
/// results. For intrinsics, we assume scalar optimizations already can fully
/// reason about them.
///
/// This repetition has the potential to be very large however, as each one
/// might refine a single call site. As a consequence, in practice we use an
/// upper bound on the number of repetitions to limit things.
template <typename PassT>
class DevirtSCCRepeatedPass
: public PassInfoMixin<DevirtSCCRepeatedPass<PassT>> {
public:
explicit DevirtSCCRepeatedPass(PassT Pass, int MaxIterations)
: Pass(std::move(Pass)), MaxIterations(MaxIterations) {}
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
/// Runs the wrapped pass up to \c MaxIterations on the SCC, iterating
/// whenever an indirect call is refined.
PreservedAnalyses run(LazyCallGraph::SCC &InitialC, CGSCCAnalysisManager &AM,
LazyCallGraph &CG, CGSCCUpdateResult &UR) {
PreservedAnalyses PA = PreservedAnalyses::all();
// 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;
// Collect value handles for all of the indirect call sites.
SmallVector<WeakTrackingVH, 8> CallHandles;
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
// 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,
SmallVectorImpl<WeakTrackingVH> &CallHandles) {
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
assert(CallHandles.empty() && "Must start with a clear set of handles.");
SmallVector<CallCount, 4> CallCounts;
for (LazyCallGraph::Node &N : C) {
CallCounts.push_back({0, 0});
CallCount &Count = CallCounts.back();
for (Instruction &I : instructions(N.getFunction()))
if (auto CS = CallSite(&I)) {
if (CS.getCalledFunction()) {
++Count.Direct;
} else {
++Count.Indirect;
CallHandles.push_back(WeakTrackingVH(&I));
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
}
}
}
return CallCounts;
};
// Populate the initial call handles and get the initial call counts.
auto CallCounts = ScanSCC(*C, CallHandles);
for (int Iteration = 0;; ++Iteration) {
PreservedAnalyses PassPA = Pass.run(*C, AM, CG, UR);
// 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!");
assert((int)CallCounts.size() == C->size() &&
"Cannot have changed the size of the SCC!");
// Check whether any of the handles were devirtualized.
auto IsDevirtualizedHandle = [&](WeakTrackingVH &CallH) {
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
if (!CallH)
return false;
auto CS = CallSite(CallH);
if (!CS)
return false;
// If the call is still indirect, leave it alone.
Function *F = CS.getCalledFunction();
if (!F)
return false;
DEBUG(dbgs() << "Found devirutalized call from "
<< CS.getParent()->getParent()->getName() << " to "
<< F->getName() << "\n");
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
// We now have a direct call where previously we had an indirect call,
// so iterate to process this devirtualization site.
return true;
};
bool Devirt = llvm::any_of(CallHandles, IsDevirtualizedHandle);
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
// 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.
CallHandles.clear();
auto NewCallCounts = ScanSCC(*C, CallHandles);
// 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)
for (int i = 0, Size = C->size(); i < Size; ++i)
if (CallCounts[i].Indirect > NewCallCounts[i].Indirect &&
CallCounts[i].Direct < NewCallCounts[i].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) {
DEBUG(dbgs() << "Found another devirtualization after hitting the max "
"number of repetitions ("
<< MaxIterations << ") on SCC: " << *C << "\n");
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
PA.intersect(std::move(PassPA));
break;
}
DEBUG(dbgs()
<< "Repeating an SCC pass after finding a devirtualization in: "
<< *C << "\n");
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
// 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;
}
private:
PassT Pass;
int MaxIterations;
};
/// A function to deduce a function pass type and wrap it in the
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
/// templated adaptor.
template <typename PassT>
DevirtSCCRepeatedPass<PassT> createDevirtSCCRepeatedPass(PassT Pass,
int MaxIterations) {
return DevirtSCCRepeatedPass<PassT>(std::move(Pass), MaxIterations);
[PM] Introduce a devirtualization iteration layer for the new PM. This is an orthogonal and separated layer instead of being embedded inside the pass manager. While it adds a small amount of complexity, it is fairly minimal and the composability and control seems worth the cost. The logic for this ends up being nicely isolated and targeted. It should be easy to experiment with different iteration strategies wrapped around the CGSCC bottom-up walk using this kind of facility. The mechanism used to track devirtualization is the simplest one I came up with. I think it handles most of the cases the existing iteration machinery handles, but I haven't done a *very* in depth analysis. It does however match the basic intended semantics, and we can tweak or tune its exact behavior incrementally as necessary. One thing that we may want to revisit is freshly building the value handle set on each iteration. While I don't think this will be a significant cost (it is strictly fewer value handles but more churn of value handes than the old call graph), it is conceivable that we'll want a somewhat more clever tracking mechanism. My hope is to layer that on as a follow up patch with data supporting any implementation complexity it adds. This code also provides for a basic count heuristic: if the number of indirect calls decreases and the number of direct calls increases for a given function in the SCC, we assume devirtualization is responsible. This matches the heuristics currently used in the legacy pass manager. Differential Revision: https://reviews.llvm.org/D23114 llvm-svn: 290665
2016-12-28 12:07:33 +01:00
}
// Clear out the debug logging macro.
#undef DEBUG_TYPE
} // end namespace llvm
#endif // LLVM_ANALYSIS_CGSCCPASSMANAGER_H