maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/Other/cgscc-observe-devirt.ll

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

107 lines
3.2 KiB
LLVM
Raw Normal View History

[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 19:07:33 +08:00
; Make sure that even without some external devirtualization iteration tool,
; the CGSCC pass manager correctly observes and re-visits SCCs that change
; structure due to devirtualization. We trigger devirtualization here with GVN
; which forwards a store through a load and to an indirect call.
;
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
; RUN: opt -aa-pipeline=basic-aa -passes='cgscc(function-attrs)' -S < %s | FileCheck %s --check-prefix=BEFORE
; RUN: opt -aa-pipeline=basic-aa -passes='cgscc(function-attrs,function(gvn))' -S < %s | FileCheck %s --check-prefix=AFTER
;
; Also check that adding an extra CGSCC pass after the function update but
; without requiring the outer manager to iterate doesn't break any invariant.
[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 19:07:33 +08:00
; RUN: opt -aa-pipeline=basic-aa -passes='cgscc(function-attrs,function(gvn),function-attrs)' -S < %s | FileCheck %s --check-prefix=AFTER
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
[funcattrs] Add the maximal set of implied attributes to definitions Have funcattrs expand all implied attributes into the IR. This expands the infrastructure from D100400, but for definitions not declarations this time. Somewhat subtly, this mostly isn't semantic. Because the accessors did the inference, any client which used the accessor was already getting the stronger result. Clients that directly checked presence of attributes (there are some), will see a stronger result now. The old behavior can end up quite confusing for two reasons: * Without this change, we have situations where function-attrs appears to fail when inferring an attribute (as seen by a human reading IR), but that consuming code will see that it should have been implied. As a human trying to sanity check test results and study IR for optimization possibilities, this is exceeding error prone and confusing. (I'll note that I wasted several hours recently because of this.) * We can have transforms which trigger without the IR appearing (on inspection) to meet the preconditions. This change doesn't prevent this from happening (as the accessors still involve multiple checks), but it should make it less frequent. I'd argue in favor of deleting the extra checks out of the accessors after this lands, but I want that in it's own review as a) it's purely stylistic, and b) I already know there's some disagreement. Once this lands, I'm also going to do a cleanup change which will delete some now redundant duplicate predicates in the inference code, but again, that deserves to be a change of it's own. Differential Revision: https://reviews.llvm.org/D100226
2021-04-17 05:03:36 +08:00
declare void @readnone() nofree nosync readnone
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
declare void @unknown()
[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 19:07:33 +08:00
; The @test1_* checks that if we refine an indirect call to a direct call and
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
; in the process change the very structure of the call graph we also revisit
; that component of the graph and do so in an up-to-date fashion.
[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 19:07:33 +08:00
; BEFORE: define void @test1_a1() {
; AFTER: define void @test1_a1() {
define void @test1_a1() {
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
%fptr = alloca void()*
[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 19:07:33 +08:00
store void()* @test1_b2, void()** %fptr
store void()* @test1_b1, void()** %fptr
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
%f = load void()*, void()** %fptr
call void %f()
ret void
}
[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 19:07:33 +08:00
; BEFORE: define void @test1_b1() {
; AFTER: define void @test1_b1() {
define void @test1_b1() {
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
call void @unknown()
[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 19:07:33 +08:00
call void @test1_a1()
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
ret void
}
[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 19:07:33 +08:00
; BEFORE: define void @test1_a2() {
; AFTER: define void @test1_a2() #0 {
define void @test1_a2() {
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
%fptr = alloca void()*
[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 19:07:33 +08:00
store void()* @test1_b1, void()** %fptr
store void()* @test1_b2, void()** %fptr
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
%f = load void()*, void()** %fptr
call void %f()
ret void
}
[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 19:07:33 +08:00
; BEFORE: define void @test1_b2() {
; AFTER: define void @test1_b2() #0 {
define void @test1_b2() {
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
call void @readnone()
[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 19:07:33 +08:00
call void @test1_a2()
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
ret void
}
[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 19:07:33 +08:00
; The @test2_* set of functions exercise a case where running function passes
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
; introduces a new post-order relationship that was not present originally and
; makes sure we walk across the SCCs in that order.
[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 19:07:33 +08:00
; CHECK: define void @test2_a() {
define void @test2_a() {
call void @test2_b1()
call void @test2_b2()
call void @test2_b3()
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
call void @unknown()
ret void
}
[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 19:07:33 +08:00
; CHECK: define void @test2_b1() #0 {
define void @test2_b1() {
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
%fptr = alloca void()*
[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 19:07:33 +08:00
store void()* @test2_a, void()** %fptr
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
store void()* @readnone, void()** %fptr
%f = load void()*, void()** %fptr
call void %f()
ret void
}
[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 19:07:33 +08:00
; CHECK: define void @test2_b2() #0 {
define void @test2_b2() {
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
%fptr = alloca void()*
[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 19:07:33 +08:00
store void()* @test2_a, void()** %fptr
store void()* @test2_b2, void()** %fptr
store void()* @test2_b3, void()** %fptr
store void()* @test2_b1, void()** %fptr
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
%f = load void()*, void()** %fptr
call void %f()
ret void
}
[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 19:07:33 +08:00
; CHECK: define void @test2_b3() #0 {
define void @test2_b3() {
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
%fptr = alloca void()*
[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 19:07:33 +08:00
store void()* @test2_a, void()** %fptr
store void()* @test2_b2, void()** %fptr
store void()* @test2_b3, void()** %fptr
store void()* @test2_b1, void()** %fptr
[PM] Introduce basic update capabilities to the new PM's CGSCC pass manager, including both plumbing and logic to handle function pass updates. There are three fundamentally tied changes here: 1) Plumbing *some* mechanism for updating the CGSCC pass manager as the CG changes while passes are running. 2) Changing the CGSCC pass manager infrastructure to have support for the underlying graph to mutate mid-pass run. 3) Actually updating the CG after function passes run. I can separate them if necessary, but I think its really useful to have them together as the needs of #3 drove #2, and that in turn drove #1. The plumbing technique is to extend the "run" method signature with extra arguments. We provide the call graph that intrinsically is available as it is the basis of the pass manager's IR units, and an output parameter that records the results of updating the call graph during an SCC passes's run. Note that "...UpdateResult" isn't a *great* name here... suggestions very welcome. I tried a pretty frustrating number of different data structures and such for the innards of the update result. Every other one failed for one reason or another. Sometimes I just couldn't keep the layers of complexity right in my head. The thing that really worked was to just directly provide access to the underlying structures used to walk the call graph so that their updates could be informed by the *particular* nature of the change to the graph. The technique for how to make the pass management infrastructure cope with mutating graphs was also something that took a really, really large number of iterations to get to a place where I was happy. Here are some of the considerations that drove the design: - We operate at three levels within the infrastructure: RefSCC, SCC, and Node. In each case, we are working bottom up and so we want to continue to iterate on the "lowest" node as the graph changes. Look at how we iterate over nodes in an SCC running function passes as those function passes mutate the CG. We continue to iterate on the "lowest" SCC, which is the one that continues to contain the function just processed. - The call graph structure re-uses SCCs (and RefSCCs) during mutation events for the *highest* entry in the resulting new subgraph, not the lowest. This means that it is necessary to continually update the current SCC or RefSCC as it shifts. This is really surprising and subtle, and took a long time for me to work out. I actually tried changing the call graph to provide the opposite behavior, and it breaks *EVERYTHING*. The graph update algorithms are really deeply tied to this particualr pattern. - When SCCs or RefSCCs are split apart and refined and we continually re-pin our processing to the bottom one in the subgraph, we need to enqueue the newly formed SCCs and RefSCCs for subsequent processing. Queuing them presents a few challenges: 1) SCCs and RefSCCs use wildly different iteration strategies at a high level. We end up needing to converge them on worklist approaches that can be extended in order to be able to handle the mutations. 2) The order of the enqueuing need to remain bottom-up post-order so that we don't get surprising order of visitation for things like the inliner. 3) We need the worklists to have set semantics so we don't duplicate things endlessly. We don't need a *persistent* set though because we always keep processing the bottom node!!!! This is super, super surprising to me and took a long time to convince myself this is correct, but I'm pretty sure it is... Once we sink down to the bottom node, we can't re-split out the same node in any way, and the postorder of the current queue is fixed and unchanging. 4) We need to make sure that the "current" SCC or RefSCC actually gets enqueued here such that we re-visit it because we continue processing a *new*, *bottom* SCC/RefSCC. - We also need the ability to *skip* SCCs and RefSCCs that get merged into a larger component. We even need the ability to skip *nodes* from an SCC that are no longer part of that SCC. This led to the design you see in the patch which uses SetVector-based worklists. The RefSCC worklist is always empty until an update occurs and is just used to handle those RefSCCs created by updates as the others don't even exist yet and are formed on-demand during the bottom-up walk. The SCC worklist is pre-populated from the RefSCC, and we push new SCCs onto it and blacklist existing SCCs on it to get the desired processing. We then *directly* update these when updating the call graph as I was never able to find a satisfactory abstraction around the update strategy. Finally, we need to compute the updates for function passes. This is mostly used as an initial customer of all the update mechanisms to drive their design to at least cover some real set of use cases. There are a bunch of interesting things that came out of doing this: - It is really nice to do this a function at a time because that function is likely hot in the cache. This means we want even the function pass adaptor to support online updates to the call graph! - To update the call graph after arbitrary function pass mutations is quite hard. We have to build a fairly comprehensive set of data structures and then process them. Fortunately, some of this code is related to the code for building the cal graph in the first place. Unfortunately, very little of it makes any sense to share because the nature of what we're doing is so very different. I've factored out the one part that made sense at least. - We need to transfer these updates into the various structures for the CGSCC pass manager. Once those were more sanely worked out, this became relatively easier. But some of those needs necessitated changes to the LazyCallGraph interface to make it significantly easier to extract the changed SCCs from an update operation. - We also need to update the CGSCC analysis manager as the shape of the graph changes. When an SCC is merged away we need to clear analyses associated with it from the analysis manager which we didn't have support for in the analysis manager infrsatructure. New SCCs are easy! But then we have the case that the original SCC has its shape changed but remains in the call graph. There we need to *invalidate* the analyses associated with it. - We also need to invalidate analyses after we *finish* processing an SCC. But the analyses we need to invalidate here are *only those for the newly updated SCC*!!! Because we only continue processing the bottom SCC, if we split SCCs apart the original one gets invalidated once when its shape changes and is not processed farther so its analyses will be correct. It is the bottom SCC which continues being processed and needs to have the "normal" invalidation done based on the preserved analyses set. All of this is mostly background and context for the changes here. Many thanks to all the reviewers who helped here. Especially Sanjoy who caught several interesting bugs in the graph algorithms, David, Sean, and others who all helped with feedback. Differential Revision: http://reviews.llvm.org/D21464 llvm-svn: 279618
2016-08-24 17:37:14 +08:00
%f = load void()*, void()** %fptr
call void %f()
ret void
}
[funcattrs] Add the maximal set of implied attributes to definitions Have funcattrs expand all implied attributes into the IR. This expands the infrastructure from D100400, but for definitions not declarations this time. Somewhat subtly, this mostly isn't semantic. Because the accessors did the inference, any client which used the accessor was already getting the stronger result. Clients that directly checked presence of attributes (there are some), will see a stronger result now. The old behavior can end up quite confusing for two reasons: * Without this change, we have situations where function-attrs appears to fail when inferring an attribute (as seen by a human reading IR), but that consuming code will see that it should have been implied. As a human trying to sanity check test results and study IR for optimization possibilities, this is exceeding error prone and confusing. (I'll note that I wasted several hours recently because of this.) * We can have transforms which trigger without the IR appearing (on inspection) to meet the preconditions. This change doesn't prevent this from happening (as the accessors still involve multiple checks), but it should make it less frequent. I'd argue in favor of deleting the extra checks out of the accessors after this lands, but I want that in it's own review as a) it's purely stylistic, and b) I already know there's some disagreement. Once this lands, I'm also going to do a cleanup change which will delete some now redundant duplicate predicates in the inference code, but again, that deserves to be a change of it's own. Differential Revision: https://reviews.llvm.org/D100226
2021-04-17 05:03:36 +08:00
; CHECK: attributes #0 = { nofree nosync readnone }