llvm-6502/lib/Analysis/LazyCallGraph.cpp

533 lines
19 KiB
C++
Raw Normal View History

[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
//===- LazyCallGraph.cpp - Analysis of a Module's call graph --------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
#include "llvm/Analysis/LazyCallGraph.h"
[LCG] Add support for building persistent and connected SCCs to the LazyCallGraph. This is the start of the whole point of this different abstraction, but it is just the initial bits. Here is a run-down of what's going on here. I'm planning to incorporate some (or all) of this into comments going forward, hopefully with better editing and wording. =] The crux of the problem with the traditional way of building SCCs is that they are ephemeral. The new pass manager however really needs the ability to associate analysis passes and results of analysis passes with SCCs in order to expose these analysis passes to the SCC passes. Making this work is kind-of the whole point of the new pass manager. =] So, when we're building SCCs for the call graph, we actually want to build persistent nodes that stick around and can be reasoned about later. We'd also like the ability to walk the SCC graph in more complex ways than just the traditional postorder traversal of the current CGSCC walk. That means that in addition to being persistent, the SCCs need to be connected into a useful graph structure. However, we still want the SCCs to be formed lazily where possible. These constraints are quite hard to satisfy with the SCC iterator. Also, using that would bypass our ability to actually add data to the nodes of the call graph to facilite implementing the Tarjan walk. So I've re-implemented things in a more direct and embedded way. This immediately makes it easy to get the persistence and connectivity correct, and it also allows leveraging the existing nodes to simplify the algorithm. I've worked somewhat to make this implementation more closely follow the traditional paper's nomenclature and strategy, although it is still a bit obtuse because it isn't recursive, using an explicit stack and a tail call instead, and it is interruptable, resuming each time we need another SCC. The other tricky bit here, and what actually took almost all the time and trials and errors I spent building this, is exactly *what* graph structure to build for the SCCs. The naive thing to build is the call graph in its newly acyclic form. I wrote about 4 versions of this which did precisely this. Inevitably, when I experimented with them across various use cases, they became incredibly awkward. It was all implementable, but it felt like a complete wrong fit. Square peg, round hole. There were two overriding aspects that pushed me in a different direction: 1) We want to discover the SCC graph in a postorder fashion. That means the root node will be the *last* node we find. Using the call-SCC DAG as the graph structure of the SCCs results in an orphaned graph until we discover a root. 2) We will eventually want to walk the SCC graph in parallel, exploring distinct sub-graphs independently, and synchronizing at merge points. This again is not helped by the call-SCC DAG structure. The structure which, quite surprisingly, ended up being completely natural to use is the *inverse* of the call-SCC DAG. We add the leaf SCCs to the graph as "roots", and have edges to the caller SCCs. Once I switched to building this structure, everything just fell into place elegantly. Aside from general cleanups (there are FIXMEs and too few comments overall) that are still needed, the other missing piece of this is support for iterating across levels of the SCC graph. These will become useful for implementing #2, but they aren't an immediate priority. Once SCCs are in good shape, I'll be working on adding mutation support for incremental updates and adding the pass manager that this analysis enables. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206581 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 10:50:32 +00:00
#include "llvm/ADT/STLExtras.h"
#include "llvm/IR/CallSite.h"
#include "llvm/IR/InstVisitor.h"
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
#include "llvm/IR/Instructions.h"
#include "llvm/IR/PassManager.h"
#include "llvm/Support/Debug.h"
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
#define DEBUG_TYPE "lcg"
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
static void findCallees(
SmallVectorImpl<Constant *> &Worklist, SmallPtrSetImpl<Constant *> &Visited,
SmallVectorImpl<PointerUnion<Function *, LazyCallGraph::Node *>> &Callees,
DenseMap<Function *, size_t> &CalleeIndexMap) {
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
while (!Worklist.empty()) {
Constant *C = Worklist.pop_back_val();
if (Function *F = dyn_cast<Function>(C)) {
// Note that we consider *any* function with a definition to be a viable
// edge. Even if the function's definition is subject to replacement by
// some other module (say, a weak definition) there may still be
// optimizations which essentially speculate based on the definition and
// a way to check that the specific definition is in fact the one being
// used. For example, this could be done by moving the weak definition to
// a strong (internal) definition and making the weak definition be an
// alias. Then a test of the address of the weak function against the new
// strong definition's address would be an effective way to determine the
// safety of optimizing a direct call edge.
if (!F->isDeclaration() &&
CalleeIndexMap.insert(std::make_pair(F, Callees.size())).second) {
DEBUG(dbgs() << " Added callable function: " << F->getName()
<< "\n");
Callees.push_back(F);
}
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
continue;
}
for (Value *Op : C->operand_values())
if (Visited.insert(cast<Constant>(Op)))
Worklist.push_back(cast<Constant>(Op));
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
}
}
[LCG] Add support for building persistent and connected SCCs to the LazyCallGraph. This is the start of the whole point of this different abstraction, but it is just the initial bits. Here is a run-down of what's going on here. I'm planning to incorporate some (or all) of this into comments going forward, hopefully with better editing and wording. =] The crux of the problem with the traditional way of building SCCs is that they are ephemeral. The new pass manager however really needs the ability to associate analysis passes and results of analysis passes with SCCs in order to expose these analysis passes to the SCC passes. Making this work is kind-of the whole point of the new pass manager. =] So, when we're building SCCs for the call graph, we actually want to build persistent nodes that stick around and can be reasoned about later. We'd also like the ability to walk the SCC graph in more complex ways than just the traditional postorder traversal of the current CGSCC walk. That means that in addition to being persistent, the SCCs need to be connected into a useful graph structure. However, we still want the SCCs to be formed lazily where possible. These constraints are quite hard to satisfy with the SCC iterator. Also, using that would bypass our ability to actually add data to the nodes of the call graph to facilite implementing the Tarjan walk. So I've re-implemented things in a more direct and embedded way. This immediately makes it easy to get the persistence and connectivity correct, and it also allows leveraging the existing nodes to simplify the algorithm. I've worked somewhat to make this implementation more closely follow the traditional paper's nomenclature and strategy, although it is still a bit obtuse because it isn't recursive, using an explicit stack and a tail call instead, and it is interruptable, resuming each time we need another SCC. The other tricky bit here, and what actually took almost all the time and trials and errors I spent building this, is exactly *what* graph structure to build for the SCCs. The naive thing to build is the call graph in its newly acyclic form. I wrote about 4 versions of this which did precisely this. Inevitably, when I experimented with them across various use cases, they became incredibly awkward. It was all implementable, but it felt like a complete wrong fit. Square peg, round hole. There were two overriding aspects that pushed me in a different direction: 1) We want to discover the SCC graph in a postorder fashion. That means the root node will be the *last* node we find. Using the call-SCC DAG as the graph structure of the SCCs results in an orphaned graph until we discover a root. 2) We will eventually want to walk the SCC graph in parallel, exploring distinct sub-graphs independently, and synchronizing at merge points. This again is not helped by the call-SCC DAG structure. The structure which, quite surprisingly, ended up being completely natural to use is the *inverse* of the call-SCC DAG. We add the leaf SCCs to the graph as "roots", and have edges to the caller SCCs. Once I switched to building this structure, everything just fell into place elegantly. Aside from general cleanups (there are FIXMEs and too few comments overall) that are still needed, the other missing piece of this is support for iterating across levels of the SCC graph. These will become useful for implementing #2, but they aren't an immediate priority. Once SCCs are in good shape, I'll be working on adding mutation support for incremental updates and adding the pass manager that this analysis enables. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206581 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 10:50:32 +00:00
LazyCallGraph::Node::Node(LazyCallGraph &G, Function &F)
: G(&G), F(F), DFSNumber(0), LowLink(0) {
DEBUG(dbgs() << " Adding functions called by '" << F.getName()
<< "' to the graph.\n");
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
SmallVector<Constant *, 16> Worklist;
SmallPtrSet<Constant *, 16> Visited;
// Find all the potential callees in this function. First walk the
// instructions and add every operand which is a constant to the worklist.
for (BasicBlock &BB : F)
for (Instruction &I : BB)
for (Value *Op : I.operand_values())
if (Constant *C = dyn_cast<Constant>(Op))
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
if (Visited.insert(C))
Worklist.push_back(C);
// We've collected all the constant (and thus potentially function or
// function containing) operands to all of the instructions in the function.
// Process them (recursively) collecting every function found.
findCallees(Worklist, Visited, Callees, CalleeIndexMap);
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
}
LazyCallGraph::LazyCallGraph(Module &M) : NextDFSNumber(0) {
DEBUG(dbgs() << "Building CG for module: " << M.getModuleIdentifier()
<< "\n");
for (Function &F : M)
if (!F.isDeclaration() && !F.hasLocalLinkage())
if (EntryIndexMap.insert(std::make_pair(&F, EntryNodes.size())).second) {
DEBUG(dbgs() << " Adding '" << F.getName()
<< "' to entry set of the graph.\n");
EntryNodes.push_back(&F);
}
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
// Now add entry nodes for functions reachable via initializers to globals.
SmallVector<Constant *, 16> Worklist;
SmallPtrSet<Constant *, 16> Visited;
for (GlobalVariable &GV : M.globals())
if (GV.hasInitializer())
if (Visited.insert(GV.getInitializer()))
Worklist.push_back(GV.getInitializer());
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
DEBUG(dbgs() << " Adding functions referenced by global initializers to the "
"entry set.\n");
findCallees(Worklist, Visited, EntryNodes, EntryIndexMap);
[LCG] Add support for building persistent and connected SCCs to the LazyCallGraph. This is the start of the whole point of this different abstraction, but it is just the initial bits. Here is a run-down of what's going on here. I'm planning to incorporate some (or all) of this into comments going forward, hopefully with better editing and wording. =] The crux of the problem with the traditional way of building SCCs is that they are ephemeral. The new pass manager however really needs the ability to associate analysis passes and results of analysis passes with SCCs in order to expose these analysis passes to the SCC passes. Making this work is kind-of the whole point of the new pass manager. =] So, when we're building SCCs for the call graph, we actually want to build persistent nodes that stick around and can be reasoned about later. We'd also like the ability to walk the SCC graph in more complex ways than just the traditional postorder traversal of the current CGSCC walk. That means that in addition to being persistent, the SCCs need to be connected into a useful graph structure. However, we still want the SCCs to be formed lazily where possible. These constraints are quite hard to satisfy with the SCC iterator. Also, using that would bypass our ability to actually add data to the nodes of the call graph to facilite implementing the Tarjan walk. So I've re-implemented things in a more direct and embedded way. This immediately makes it easy to get the persistence and connectivity correct, and it also allows leveraging the existing nodes to simplify the algorithm. I've worked somewhat to make this implementation more closely follow the traditional paper's nomenclature and strategy, although it is still a bit obtuse because it isn't recursive, using an explicit stack and a tail call instead, and it is interruptable, resuming each time we need another SCC. The other tricky bit here, and what actually took almost all the time and trials and errors I spent building this, is exactly *what* graph structure to build for the SCCs. The naive thing to build is the call graph in its newly acyclic form. I wrote about 4 versions of this which did precisely this. Inevitably, when I experimented with them across various use cases, they became incredibly awkward. It was all implementable, but it felt like a complete wrong fit. Square peg, round hole. There were two overriding aspects that pushed me in a different direction: 1) We want to discover the SCC graph in a postorder fashion. That means the root node will be the *last* node we find. Using the call-SCC DAG as the graph structure of the SCCs results in an orphaned graph until we discover a root. 2) We will eventually want to walk the SCC graph in parallel, exploring distinct sub-graphs independently, and synchronizing at merge points. This again is not helped by the call-SCC DAG structure. The structure which, quite surprisingly, ended up being completely natural to use is the *inverse* of the call-SCC DAG. We add the leaf SCCs to the graph as "roots", and have edges to the caller SCCs. Once I switched to building this structure, everything just fell into place elegantly. Aside from general cleanups (there are FIXMEs and too few comments overall) that are still needed, the other missing piece of this is support for iterating across levels of the SCC graph. These will become useful for implementing #2, but they aren't an immediate priority. Once SCCs are in good shape, I'll be working on adding mutation support for incremental updates and adding the pass manager that this analysis enables. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206581 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 10:50:32 +00:00
for (auto &Entry : EntryNodes)
if (Function *F = Entry.dyn_cast<Function *>())
SCCEntryNodes.insert(F);
else
SCCEntryNodes.insert(&Entry.get<Node *>()->getFunction());
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
}
LazyCallGraph::LazyCallGraph(LazyCallGraph &&G)
: BPA(std::move(G.BPA)), NodeMap(std::move(G.NodeMap)),
EntryNodes(std::move(G.EntryNodes)),
EntryIndexMap(std::move(G.EntryIndexMap)), SCCBPA(std::move(G.SCCBPA)),
[LCG] Add support for building persistent and connected SCCs to the LazyCallGraph. This is the start of the whole point of this different abstraction, but it is just the initial bits. Here is a run-down of what's going on here. I'm planning to incorporate some (or all) of this into comments going forward, hopefully with better editing and wording. =] The crux of the problem with the traditional way of building SCCs is that they are ephemeral. The new pass manager however really needs the ability to associate analysis passes and results of analysis passes with SCCs in order to expose these analysis passes to the SCC passes. Making this work is kind-of the whole point of the new pass manager. =] So, when we're building SCCs for the call graph, we actually want to build persistent nodes that stick around and can be reasoned about later. We'd also like the ability to walk the SCC graph in more complex ways than just the traditional postorder traversal of the current CGSCC walk. That means that in addition to being persistent, the SCCs need to be connected into a useful graph structure. However, we still want the SCCs to be formed lazily where possible. These constraints are quite hard to satisfy with the SCC iterator. Also, using that would bypass our ability to actually add data to the nodes of the call graph to facilite implementing the Tarjan walk. So I've re-implemented things in a more direct and embedded way. This immediately makes it easy to get the persistence and connectivity correct, and it also allows leveraging the existing nodes to simplify the algorithm. I've worked somewhat to make this implementation more closely follow the traditional paper's nomenclature and strategy, although it is still a bit obtuse because it isn't recursive, using an explicit stack and a tail call instead, and it is interruptable, resuming each time we need another SCC. The other tricky bit here, and what actually took almost all the time and trials and errors I spent building this, is exactly *what* graph structure to build for the SCCs. The naive thing to build is the call graph in its newly acyclic form. I wrote about 4 versions of this which did precisely this. Inevitably, when I experimented with them across various use cases, they became incredibly awkward. It was all implementable, but it felt like a complete wrong fit. Square peg, round hole. There were two overriding aspects that pushed me in a different direction: 1) We want to discover the SCC graph in a postorder fashion. That means the root node will be the *last* node we find. Using the call-SCC DAG as the graph structure of the SCCs results in an orphaned graph until we discover a root. 2) We will eventually want to walk the SCC graph in parallel, exploring distinct sub-graphs independently, and synchronizing at merge points. This again is not helped by the call-SCC DAG structure. The structure which, quite surprisingly, ended up being completely natural to use is the *inverse* of the call-SCC DAG. We add the leaf SCCs to the graph as "roots", and have edges to the caller SCCs. Once I switched to building this structure, everything just fell into place elegantly. Aside from general cleanups (there are FIXMEs and too few comments overall) that are still needed, the other missing piece of this is support for iterating across levels of the SCC graph. These will become useful for implementing #2, but they aren't an immediate priority. Once SCCs are in good shape, I'll be working on adding mutation support for incremental updates and adding the pass manager that this analysis enables. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206581 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 10:50:32 +00:00
SCCMap(std::move(G.SCCMap)), LeafSCCs(std::move(G.LeafSCCs)),
DFSStack(std::move(G.DFSStack)),
SCCEntryNodes(std::move(G.SCCEntryNodes)),
NextDFSNumber(G.NextDFSNumber) {
updateGraphPtrs();
}
LazyCallGraph &LazyCallGraph::operator=(LazyCallGraph &&G) {
BPA = std::move(G.BPA);
NodeMap = std::move(G.NodeMap);
EntryNodes = std::move(G.EntryNodes);
EntryIndexMap = std::move(G.EntryIndexMap);
SCCBPA = std::move(G.SCCBPA);
SCCMap = std::move(G.SCCMap);
LeafSCCs = std::move(G.LeafSCCs);
DFSStack = std::move(G.DFSStack);
SCCEntryNodes = std::move(G.SCCEntryNodes);
NextDFSNumber = G.NextDFSNumber;
updateGraphPtrs();
return *this;
}
[LCG] Add the first round of mutation support to the lazy call graph. This implements the core functionality necessary to remove an edge from the call graph and correctly update both the basic graph and the SCC structure. As part of that it has to run a tiny (in number of nodes) Tarjan-style DFS walk of an SCC being mutated to compute newly formed SCCs, etc. This is *very rough* and a WIP. I have a bunch of FIXMEs for code cleanup that will reduce the boilerplate in this change substantially. I also have a bunch of simplifications to various parts of both algorithms that I want to make, but first I'd like to have a more holistic picture. Ideally, I'd also like more testing. I'll probably add quite a few more unit tests as I go here to cover the various different aspects and corner cases of removing edges from the graph. Still, this is, so far, successfully updating the SCC graph in-place without disrupting the identity established for the existing SCCs even when we do challenging things like delete the critical edge that made an SCC cycle at all and have to reform things as a tree of smaller SCCs. Getting this to work is really critical for the new pass manager as it is going to associate significant state with the SCC instance and needs it to be stable. That is also the motivation behind the return of the newly formed SCCs. Eventually, I'll wire this all the way up to the public API so that the pass manager can use it to correctly re-enqueue newly formed SCCs into a fresh postorder traversal. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206968 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-23 11:03:03 +00:00
void LazyCallGraph::SCC::removeEdge(LazyCallGraph &G, Function &Caller,
Function &Callee, SCC &CalleeC) {
assert(std::find(G.LeafSCCs.begin(), G.LeafSCCs.end(), this) ==
G.LeafSCCs.end() &&
"Cannot have a leaf SCC caller with a different SCC callee.");
bool HasOtherCallToCalleeC = false;
bool HasOtherCallOutsideSCC = false;
for (Node *N : *this) {
for (Node *Callee : *N) {
SCC *OtherCalleeC = G.SCCMap.lookup(Callee);
[LCG] Add the first round of mutation support to the lazy call graph. This implements the core functionality necessary to remove an edge from the call graph and correctly update both the basic graph and the SCC structure. As part of that it has to run a tiny (in number of nodes) Tarjan-style DFS walk of an SCC being mutated to compute newly formed SCCs, etc. This is *very rough* and a WIP. I have a bunch of FIXMEs for code cleanup that will reduce the boilerplate in this change substantially. I also have a bunch of simplifications to various parts of both algorithms that I want to make, but first I'd like to have a more holistic picture. Ideally, I'd also like more testing. I'll probably add quite a few more unit tests as I go here to cover the various different aspects and corner cases of removing edges from the graph. Still, this is, so far, successfully updating the SCC graph in-place without disrupting the identity established for the existing SCCs even when we do challenging things like delete the critical edge that made an SCC cycle at all and have to reform things as a tree of smaller SCCs. Getting this to work is really critical for the new pass manager as it is going to associate significant state with the SCC instance and needs it to be stable. That is also the motivation behind the return of the newly formed SCCs. Eventually, I'll wire this all the way up to the public API so that the pass manager can use it to correctly re-enqueue newly formed SCCs into a fresh postorder traversal. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206968 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-23 11:03:03 +00:00
if (OtherCalleeC == &CalleeC) {
HasOtherCallToCalleeC = true;
break;
}
if (OtherCalleeC != this)
HasOtherCallOutsideSCC = true;
}
if (HasOtherCallToCalleeC)
break;
}
// Because the SCCs form a DAG, deleting such an edge cannot change the set
// of SCCs in the graph. However, it may cut an edge of the SCC DAG, making
// the caller no longer a parent of the callee. Walk the other call edges
// in the caller to tell.
if (!HasOtherCallToCalleeC) {
bool Removed = CalleeC.ParentSCCs.remove(this);
(void)Removed;
assert(Removed &&
"Did not find the caller SCC in the callee SCC's parent list!");
// It may orphan an SCC if it is the last edge reaching it, but that does
// not violate any invariants of the graph.
if (CalleeC.ParentSCCs.empty())
DEBUG(dbgs() << "LCG: Update removing " << Caller.getName() << " -> "
<< Callee.getName() << " edge orphaned the callee's SCC!\n");
}
// It may make the Caller SCC a leaf SCC.
if (!HasOtherCallOutsideSCC)
G.LeafSCCs.push_back(this);
}
SmallVector<LazyCallGraph::SCC *, 1>
LazyCallGraph::SCC::removeInternalEdge(LazyCallGraph &G, Node &Caller,
Node &Callee) {
// We return a list of the resulting SCCs, where 'this' is always the first
// element.
SmallVector<SCC *, 1> ResultSCCs;
ResultSCCs.push_back(this);
// We're going to do a full mini-Tarjan's walk using a local stack here.
int NextDFSNumber = 1;
SmallVector<std::pair<Node *, Node::iterator>, 4> DFSStack;
// The worklist is every node in the original SCC. FIXME: switch the SCC to
// use a SmallSetVector and swap here.
SmallSetVector<Node *, 1> Worklist;
for (Node *N : Nodes) {
// Clear these to 0 while we re-run Tarjan's over the SCC.
N->DFSNumber = 0;
N->LowLink = 0;
Worklist.insert(N);
}
// The callee can already reach every node in this SCC (by definition). It is
// the only node we know will stay inside this SCC. Everything which
// transitively reaches Callee will also remain in the SCC. To model this we
// incrementally add any chain of nodes which reaches something in the new
// node set to the new node set. This short circuits one side of the Tarjan's
// walk.
SmallSetVector<Node *, 1> NewNodes;
NewNodes.insert(&Callee);
for (;;) {
if (DFSStack.empty()) {
if (Worklist.empty())
break;
Node *N = Worklist.pop_back_val();
DFSStack.push_back(std::make_pair(N, N->begin()));
}
Node *N = DFSStack.back().first;
// Check if we have reached a node in the new (known connected) set. If so,
// the entire stack is necessarily in that set and we can re-start.
if (NewNodes.count(N)) {
DFSStack.pop_back();
while (!DFSStack.empty())
NewNodes.insert(DFSStack.pop_back_val().first);
continue;
}
if (N->DFSNumber == 0) {
N->LowLink = N->DFSNumber = NextDFSNumber++;
Worklist.remove(N);
}
auto SI = DFSStack.rbegin();
bool PushedChildNode = false;
do {
N = SI->first;
for (auto I = SI->second, E = N->end(); I != E; ++I) {
Node *ChildN = *I;
// If this child isn't currently in this SCC, no need to process it.
// However, we do need to remove this SCC from its SCC's parent set.
SCC *ChildSCC = G.SCCMap.lookup(ChildN);
[LCG] Add the first round of mutation support to the lazy call graph. This implements the core functionality necessary to remove an edge from the call graph and correctly update both the basic graph and the SCC structure. As part of that it has to run a tiny (in number of nodes) Tarjan-style DFS walk of an SCC being mutated to compute newly formed SCCs, etc. This is *very rough* and a WIP. I have a bunch of FIXMEs for code cleanup that will reduce the boilerplate in this change substantially. I also have a bunch of simplifications to various parts of both algorithms that I want to make, but first I'd like to have a more holistic picture. Ideally, I'd also like more testing. I'll probably add quite a few more unit tests as I go here to cover the various different aspects and corner cases of removing edges from the graph. Still, this is, so far, successfully updating the SCC graph in-place without disrupting the identity established for the existing SCCs even when we do challenging things like delete the critical edge that made an SCC cycle at all and have to reform things as a tree of smaller SCCs. Getting this to work is really critical for the new pass manager as it is going to associate significant state with the SCC instance and needs it to be stable. That is also the motivation behind the return of the newly formed SCCs. Eventually, I'll wire this all the way up to the public API so that the pass manager can use it to correctly re-enqueue newly formed SCCs into a fresh postorder traversal. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206968 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-23 11:03:03 +00:00
assert(ChildSCC &&
"Everything reachable must already be in *some* SCC");
if (ChildSCC != this) {
ChildSCC->ParentSCCs.remove(this);
continue;
}
if (ChildN->DFSNumber == 0) {
// Mark that we should start at this child when next this node is the
// top of the stack. We don't start at the next child to ensure this
// child's lowlink is reflected.
SI->second = I;
// Recurse onto this node via a tail call.
DFSStack.push_back(std::make_pair(ChildN, ChildN->begin()));
PushedChildNode = true;
break;
}
// Track the lowest link of the childen, if any are still in the stack.
// Any child not on the stack will have a LowLink of -1.
assert(ChildN->LowLink != 0 &&
"Low-link must not be zero with a non-zero DFS number.");
[LCG] Add the first round of mutation support to the lazy call graph. This implements the core functionality necessary to remove an edge from the call graph and correctly update both the basic graph and the SCC structure. As part of that it has to run a tiny (in number of nodes) Tarjan-style DFS walk of an SCC being mutated to compute newly formed SCCs, etc. This is *very rough* and a WIP. I have a bunch of FIXMEs for code cleanup that will reduce the boilerplate in this change substantially. I also have a bunch of simplifications to various parts of both algorithms that I want to make, but first I'd like to have a more holistic picture. Ideally, I'd also like more testing. I'll probably add quite a few more unit tests as I go here to cover the various different aspects and corner cases of removing edges from the graph. Still, this is, so far, successfully updating the SCC graph in-place without disrupting the identity established for the existing SCCs even when we do challenging things like delete the critical edge that made an SCC cycle at all and have to reform things as a tree of smaller SCCs. Getting this to work is really critical for the new pass manager as it is going to associate significant state with the SCC instance and needs it to be stable. That is also the motivation behind the return of the newly formed SCCs. Eventually, I'll wire this all the way up to the public API so that the pass manager can use it to correctly re-enqueue newly formed SCCs into a fresh postorder traversal. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206968 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-23 11:03:03 +00:00
if (ChildN->LowLink >= 0 && ChildN->LowLink < N->LowLink)
N->LowLink = ChildN->LowLink;
}
if (!PushedChildNode)
// No more children to process for this stack entry.
SI->second = N->end();
++SI;
// If nothing is new on the stack and this isn't the SCC root, walk
// upward.
} while (!PushedChildNode && N->LowLink != N->DFSNumber &&
SI != DFSStack.rend());
if (PushedChildNode)
continue;
// Form the new SCC out of the top of the DFS stack.
ResultSCCs.push_back(G.formSCCFromDFSStack(DFSStack, SI.base()));
}
// Replace this SCC with the NewNodes we collected above.
// FIXME: Simplify this when the SCC's datastructure is just a list.
Nodes.clear();
NodeSet.clear();
// Now we need to reconnect the current SCC to the graph.
bool IsLeafSCC = true;
for (Node *N : NewNodes) {
N->DFSNumber = -1;
N->LowLink = -1;
Nodes.push_back(N);
NodeSet.insert(&N->getFunction());
for (Node *ChildN : *N) {
if (NewNodes.count(ChildN))
continue;
SCC *ChildSCC = G.SCCMap.lookup(ChildN);
[LCG] Add the first round of mutation support to the lazy call graph. This implements the core functionality necessary to remove an edge from the call graph and correctly update both the basic graph and the SCC structure. As part of that it has to run a tiny (in number of nodes) Tarjan-style DFS walk of an SCC being mutated to compute newly formed SCCs, etc. This is *very rough* and a WIP. I have a bunch of FIXMEs for code cleanup that will reduce the boilerplate in this change substantially. I also have a bunch of simplifications to various parts of both algorithms that I want to make, but first I'd like to have a more holistic picture. Ideally, I'd also like more testing. I'll probably add quite a few more unit tests as I go here to cover the various different aspects and corner cases of removing edges from the graph. Still, this is, so far, successfully updating the SCC graph in-place without disrupting the identity established for the existing SCCs even when we do challenging things like delete the critical edge that made an SCC cycle at all and have to reform things as a tree of smaller SCCs. Getting this to work is really critical for the new pass manager as it is going to associate significant state with the SCC instance and needs it to be stable. That is also the motivation behind the return of the newly formed SCCs. Eventually, I'll wire this all the way up to the public API so that the pass manager can use it to correctly re-enqueue newly formed SCCs into a fresh postorder traversal. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206968 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-23 11:03:03 +00:00
assert(ChildSCC &&
"Must have all child SCCs processed when building a new SCC!");
ChildSCC->ParentSCCs.insert(this);
IsLeafSCC = false;
}
}
#ifndef NDEBUG
if (ResultSCCs.size() > 1)
assert(!IsLeafSCC && "This SCC cannot be a leaf as we have split out new "
"SCCs by removing this edge.");
if (!std::any_of(G.LeafSCCs.begin(), G.LeafSCCs.end(),
[&](SCC *C) { return C == this; }))
assert(!IsLeafSCC && "This SCC cannot be a leaf as it already had child "
"SCCs before we removed this edge.");
#endif
// If this SCC stopped being a leaf through this edge removal, remove it from
// the leaf SCC list.
if (!IsLeafSCC && ResultSCCs.size() > 1)
G.LeafSCCs.erase(std::remove(G.LeafSCCs.begin(), G.LeafSCCs.end(), this),
G.LeafSCCs.end());
// Return the new list of SCCs.
return ResultSCCs;
}
void LazyCallGraph::removeEdge(Node &CallerN, Function &Callee) {
auto IndexMapI = CallerN.CalleeIndexMap.find(&Callee);
assert(IndexMapI != CallerN.CalleeIndexMap.end() &&
"Callee not in the callee set for the caller?");
Node *CalleeN = CallerN.Callees[IndexMapI->second].dyn_cast<Node *>();
CallerN.Callees.erase(CallerN.Callees.begin() + IndexMapI->second);
CallerN.CalleeIndexMap.erase(IndexMapI);
SCC *CallerC = SCCMap.lookup(&CallerN);
[LCG] Add the first round of mutation support to the lazy call graph. This implements the core functionality necessary to remove an edge from the call graph and correctly update both the basic graph and the SCC structure. As part of that it has to run a tiny (in number of nodes) Tarjan-style DFS walk of an SCC being mutated to compute newly formed SCCs, etc. This is *very rough* and a WIP. I have a bunch of FIXMEs for code cleanup that will reduce the boilerplate in this change substantially. I also have a bunch of simplifications to various parts of both algorithms that I want to make, but first I'd like to have a more holistic picture. Ideally, I'd also like more testing. I'll probably add quite a few more unit tests as I go here to cover the various different aspects and corner cases of removing edges from the graph. Still, this is, so far, successfully updating the SCC graph in-place without disrupting the identity established for the existing SCCs even when we do challenging things like delete the critical edge that made an SCC cycle at all and have to reform things as a tree of smaller SCCs. Getting this to work is really critical for the new pass manager as it is going to associate significant state with the SCC instance and needs it to be stable. That is also the motivation behind the return of the newly formed SCCs. Eventually, I'll wire this all the way up to the public API so that the pass manager can use it to correctly re-enqueue newly formed SCCs into a fresh postorder traversal. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206968 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-23 11:03:03 +00:00
if (!CallerC) {
// We can only remove edges when the edge isn't actively participating in
// a DFS walk. Either it must have been popped into an SCC, or it must not
// yet have been reached by the DFS walk. Assert the latter here.
assert(std::all_of(DFSStack.begin(), DFSStack.end(),
[&](const std::pair<Node *, iterator> &StackEntry) {
return StackEntry.first != &CallerN;
}) &&
"Found the caller on the DFSStack!");
return;
}
assert(CalleeN && "If the caller is in an SCC, we have to have explored all "
"its transitively called functions.");
SCC *CalleeC = SCCMap.lookup(CalleeN);
[LCG] Add the first round of mutation support to the lazy call graph. This implements the core functionality necessary to remove an edge from the call graph and correctly update both the basic graph and the SCC structure. As part of that it has to run a tiny (in number of nodes) Tarjan-style DFS walk of an SCC being mutated to compute newly formed SCCs, etc. This is *very rough* and a WIP. I have a bunch of FIXMEs for code cleanup that will reduce the boilerplate in this change substantially. I also have a bunch of simplifications to various parts of both algorithms that I want to make, but first I'd like to have a more holistic picture. Ideally, I'd also like more testing. I'll probably add quite a few more unit tests as I go here to cover the various different aspects and corner cases of removing edges from the graph. Still, this is, so far, successfully updating the SCC graph in-place without disrupting the identity established for the existing SCCs even when we do challenging things like delete the critical edge that made an SCC cycle at all and have to reform things as a tree of smaller SCCs. Getting this to work is really critical for the new pass manager as it is going to associate significant state with the SCC instance and needs it to be stable. That is also the motivation behind the return of the newly formed SCCs. Eventually, I'll wire this all the way up to the public API so that the pass manager can use it to correctly re-enqueue newly formed SCCs into a fresh postorder traversal. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206968 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-23 11:03:03 +00:00
assert(CalleeC &&
"The caller has an SCC, and thus by necessity so does the callee.");
// The easy case is when they are different SCCs.
if (CallerC != CalleeC) {
CallerC->removeEdge(*this, CallerN.getFunction(), Callee, *CalleeC);
return;
}
// The hard case is when we remove an edge within a SCC. This may cause new
// SCCs to need to be added to the graph.
CallerC->removeInternalEdge(*this, CallerN, *CalleeN);
}
LazyCallGraph::Node &LazyCallGraph::insertInto(Function &F, Node *&MappedN) {
return *new (MappedN = BPA.Allocate()) Node(*this, F);
}
void LazyCallGraph::updateGraphPtrs() {
// Process all nodes updating the graph pointers.
SmallVector<Node *, 16> Worklist;
for (auto &Entry : EntryNodes)
if (Node *EntryN = Entry.dyn_cast<Node *>())
Worklist.push_back(EntryN);
while (!Worklist.empty()) {
Node *N = Worklist.pop_back_val();
N->G = this;
for (auto &Callee : N->Callees)
if (Node *CalleeN = Callee.dyn_cast<Node *>())
Worklist.push_back(CalleeN);
}
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
}
LazyCallGraph::SCC *LazyCallGraph::formSCCFromDFSStack(
SmallVectorImpl<std::pair<Node *, Node::iterator>> &DFSStack,
SmallVectorImpl<std::pair<Node *, Node::iterator>>::iterator SCCBegin) {
[LCG] Add support for building persistent and connected SCCs to the LazyCallGraph. This is the start of the whole point of this different abstraction, but it is just the initial bits. Here is a run-down of what's going on here. I'm planning to incorporate some (or all) of this into comments going forward, hopefully with better editing and wording. =] The crux of the problem with the traditional way of building SCCs is that they are ephemeral. The new pass manager however really needs the ability to associate analysis passes and results of analysis passes with SCCs in order to expose these analysis passes to the SCC passes. Making this work is kind-of the whole point of the new pass manager. =] So, when we're building SCCs for the call graph, we actually want to build persistent nodes that stick around and can be reasoned about later. We'd also like the ability to walk the SCC graph in more complex ways than just the traditional postorder traversal of the current CGSCC walk. That means that in addition to being persistent, the SCCs need to be connected into a useful graph structure. However, we still want the SCCs to be formed lazily where possible. These constraints are quite hard to satisfy with the SCC iterator. Also, using that would bypass our ability to actually add data to the nodes of the call graph to facilite implementing the Tarjan walk. So I've re-implemented things in a more direct and embedded way. This immediately makes it easy to get the persistence and connectivity correct, and it also allows leveraging the existing nodes to simplify the algorithm. I've worked somewhat to make this implementation more closely follow the traditional paper's nomenclature and strategy, although it is still a bit obtuse because it isn't recursive, using an explicit stack and a tail call instead, and it is interruptable, resuming each time we need another SCC. The other tricky bit here, and what actually took almost all the time and trials and errors I spent building this, is exactly *what* graph structure to build for the SCCs. The naive thing to build is the call graph in its newly acyclic form. I wrote about 4 versions of this which did precisely this. Inevitably, when I experimented with them across various use cases, they became incredibly awkward. It was all implementable, but it felt like a complete wrong fit. Square peg, round hole. There were two overriding aspects that pushed me in a different direction: 1) We want to discover the SCC graph in a postorder fashion. That means the root node will be the *last* node we find. Using the call-SCC DAG as the graph structure of the SCCs results in an orphaned graph until we discover a root. 2) We will eventually want to walk the SCC graph in parallel, exploring distinct sub-graphs independently, and synchronizing at merge points. This again is not helped by the call-SCC DAG structure. The structure which, quite surprisingly, ended up being completely natural to use is the *inverse* of the call-SCC DAG. We add the leaf SCCs to the graph as "roots", and have edges to the caller SCCs. Once I switched to building this structure, everything just fell into place elegantly. Aside from general cleanups (there are FIXMEs and too few comments overall) that are still needed, the other missing piece of this is support for iterating across levels of the SCC graph. These will become useful for implementing #2, but they aren't an immediate priority. Once SCCs are in good shape, I'll be working on adding mutation support for incremental updates and adding the pass manager that this analysis enables. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206581 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 10:50:32 +00:00
// The tail of the stack is the new SCC. Allocate the SCC and pop the stack
// into it.
SCC *NewSCC = new (SCCBPA.Allocate()) SCC();
for (auto I = SCCBegin, E = DFSStack.end(); I != E; ++I) {
Node *SCCN = I->first;
assert(SCCN->LowLink >= SCCBegin->first->LowLink &&
"We cannot have a low link in an SCC lower than its root on the "
"stack!");
SCCMap[SCCN] = NewSCC;
[LCG] Add support for building persistent and connected SCCs to the LazyCallGraph. This is the start of the whole point of this different abstraction, but it is just the initial bits. Here is a run-down of what's going on here. I'm planning to incorporate some (or all) of this into comments going forward, hopefully with better editing and wording. =] The crux of the problem with the traditional way of building SCCs is that they are ephemeral. The new pass manager however really needs the ability to associate analysis passes and results of analysis passes with SCCs in order to expose these analysis passes to the SCC passes. Making this work is kind-of the whole point of the new pass manager. =] So, when we're building SCCs for the call graph, we actually want to build persistent nodes that stick around and can be reasoned about later. We'd also like the ability to walk the SCC graph in more complex ways than just the traditional postorder traversal of the current CGSCC walk. That means that in addition to being persistent, the SCCs need to be connected into a useful graph structure. However, we still want the SCCs to be formed lazily where possible. These constraints are quite hard to satisfy with the SCC iterator. Also, using that would bypass our ability to actually add data to the nodes of the call graph to facilite implementing the Tarjan walk. So I've re-implemented things in a more direct and embedded way. This immediately makes it easy to get the persistence and connectivity correct, and it also allows leveraging the existing nodes to simplify the algorithm. I've worked somewhat to make this implementation more closely follow the traditional paper's nomenclature and strategy, although it is still a bit obtuse because it isn't recursive, using an explicit stack and a tail call instead, and it is interruptable, resuming each time we need another SCC. The other tricky bit here, and what actually took almost all the time and trials and errors I spent building this, is exactly *what* graph structure to build for the SCCs. The naive thing to build is the call graph in its newly acyclic form. I wrote about 4 versions of this which did precisely this. Inevitably, when I experimented with them across various use cases, they became incredibly awkward. It was all implementable, but it felt like a complete wrong fit. Square peg, round hole. There were two overriding aspects that pushed me in a different direction: 1) We want to discover the SCC graph in a postorder fashion. That means the root node will be the *last* node we find. Using the call-SCC DAG as the graph structure of the SCCs results in an orphaned graph until we discover a root. 2) We will eventually want to walk the SCC graph in parallel, exploring distinct sub-graphs independently, and synchronizing at merge points. This again is not helped by the call-SCC DAG structure. The structure which, quite surprisingly, ended up being completely natural to use is the *inverse* of the call-SCC DAG. We add the leaf SCCs to the graph as "roots", and have edges to the caller SCCs. Once I switched to building this structure, everything just fell into place elegantly. Aside from general cleanups (there are FIXMEs and too few comments overall) that are still needed, the other missing piece of this is support for iterating across levels of the SCC graph. These will become useful for implementing #2, but they aren't an immediate priority. Once SCCs are in good shape, I'll be working on adding mutation support for incremental updates and adding the pass manager that this analysis enables. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206581 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 10:50:32 +00:00
NewSCC->Nodes.push_back(SCCN);
bool Inserted =
NewSCC->NodeSet.insert(&SCCN->getFunction());
(void)Inserted;
assert(Inserted && "Cannot have duplicates in the DFSStack!");
}
DFSStack.erase(SCCBegin, DFSStack.end());
[LCG] Add support for building persistent and connected SCCs to the LazyCallGraph. This is the start of the whole point of this different abstraction, but it is just the initial bits. Here is a run-down of what's going on here. I'm planning to incorporate some (or all) of this into comments going forward, hopefully with better editing and wording. =] The crux of the problem with the traditional way of building SCCs is that they are ephemeral. The new pass manager however really needs the ability to associate analysis passes and results of analysis passes with SCCs in order to expose these analysis passes to the SCC passes. Making this work is kind-of the whole point of the new pass manager. =] So, when we're building SCCs for the call graph, we actually want to build persistent nodes that stick around and can be reasoned about later. We'd also like the ability to walk the SCC graph in more complex ways than just the traditional postorder traversal of the current CGSCC walk. That means that in addition to being persistent, the SCCs need to be connected into a useful graph structure. However, we still want the SCCs to be formed lazily where possible. These constraints are quite hard to satisfy with the SCC iterator. Also, using that would bypass our ability to actually add data to the nodes of the call graph to facilite implementing the Tarjan walk. So I've re-implemented things in a more direct and embedded way. This immediately makes it easy to get the persistence and connectivity correct, and it also allows leveraging the existing nodes to simplify the algorithm. I've worked somewhat to make this implementation more closely follow the traditional paper's nomenclature and strategy, although it is still a bit obtuse because it isn't recursive, using an explicit stack and a tail call instead, and it is interruptable, resuming each time we need another SCC. The other tricky bit here, and what actually took almost all the time and trials and errors I spent building this, is exactly *what* graph structure to build for the SCCs. The naive thing to build is the call graph in its newly acyclic form. I wrote about 4 versions of this which did precisely this. Inevitably, when I experimented with them across various use cases, they became incredibly awkward. It was all implementable, but it felt like a complete wrong fit. Square peg, round hole. There were two overriding aspects that pushed me in a different direction: 1) We want to discover the SCC graph in a postorder fashion. That means the root node will be the *last* node we find. Using the call-SCC DAG as the graph structure of the SCCs results in an orphaned graph until we discover a root. 2) We will eventually want to walk the SCC graph in parallel, exploring distinct sub-graphs independently, and synchronizing at merge points. This again is not helped by the call-SCC DAG structure. The structure which, quite surprisingly, ended up being completely natural to use is the *inverse* of the call-SCC DAG. We add the leaf SCCs to the graph as "roots", and have edges to the caller SCCs. Once I switched to building this structure, everything just fell into place elegantly. Aside from general cleanups (there are FIXMEs and too few comments overall) that are still needed, the other missing piece of this is support for iterating across levels of the SCC graph. These will become useful for implementing #2, but they aren't an immediate priority. Once SCCs are in good shape, I'll be working on adding mutation support for incremental updates and adding the pass manager that this analysis enables. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206581 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 10:50:32 +00:00
// A final pass over all edges in the SCC (this remains linear as we only
// do this once when we build the SCC) to connect it to the parent sets of
// its children.
bool IsLeafSCC = true;
for (Node *SCCN : NewSCC->Nodes)
for (Node *SCCChildN : *SCCN) {
if (NewSCC->NodeSet.count(&SCCChildN->getFunction()))
continue;
SCC *ChildSCC = SCCMap.lookup(SCCChildN);
[LCG] Add support for building persistent and connected SCCs to the LazyCallGraph. This is the start of the whole point of this different abstraction, but it is just the initial bits. Here is a run-down of what's going on here. I'm planning to incorporate some (or all) of this into comments going forward, hopefully with better editing and wording. =] The crux of the problem with the traditional way of building SCCs is that they are ephemeral. The new pass manager however really needs the ability to associate analysis passes and results of analysis passes with SCCs in order to expose these analysis passes to the SCC passes. Making this work is kind-of the whole point of the new pass manager. =] So, when we're building SCCs for the call graph, we actually want to build persistent nodes that stick around and can be reasoned about later. We'd also like the ability to walk the SCC graph in more complex ways than just the traditional postorder traversal of the current CGSCC walk. That means that in addition to being persistent, the SCCs need to be connected into a useful graph structure. However, we still want the SCCs to be formed lazily where possible. These constraints are quite hard to satisfy with the SCC iterator. Also, using that would bypass our ability to actually add data to the nodes of the call graph to facilite implementing the Tarjan walk. So I've re-implemented things in a more direct and embedded way. This immediately makes it easy to get the persistence and connectivity correct, and it also allows leveraging the existing nodes to simplify the algorithm. I've worked somewhat to make this implementation more closely follow the traditional paper's nomenclature and strategy, although it is still a bit obtuse because it isn't recursive, using an explicit stack and a tail call instead, and it is interruptable, resuming each time we need another SCC. The other tricky bit here, and what actually took almost all the time and trials and errors I spent building this, is exactly *what* graph structure to build for the SCCs. The naive thing to build is the call graph in its newly acyclic form. I wrote about 4 versions of this which did precisely this. Inevitably, when I experimented with them across various use cases, they became incredibly awkward. It was all implementable, but it felt like a complete wrong fit. Square peg, round hole. There were two overriding aspects that pushed me in a different direction: 1) We want to discover the SCC graph in a postorder fashion. That means the root node will be the *last* node we find. Using the call-SCC DAG as the graph structure of the SCCs results in an orphaned graph until we discover a root. 2) We will eventually want to walk the SCC graph in parallel, exploring distinct sub-graphs independently, and synchronizing at merge points. This again is not helped by the call-SCC DAG structure. The structure which, quite surprisingly, ended up being completely natural to use is the *inverse* of the call-SCC DAG. We add the leaf SCCs to the graph as "roots", and have edges to the caller SCCs. Once I switched to building this structure, everything just fell into place elegantly. Aside from general cleanups (there are FIXMEs and too few comments overall) that are still needed, the other missing piece of this is support for iterating across levels of the SCC graph. These will become useful for implementing #2, but they aren't an immediate priority. Once SCCs are in good shape, I'll be working on adding mutation support for incremental updates and adding the pass manager that this analysis enables. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206581 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 10:50:32 +00:00
assert(ChildSCC &&
"Must have all child SCCs processed when building a new SCC!");
ChildSCC->ParentSCCs.insert(NewSCC);
IsLeafSCC = false;
}
// For the SCCs where we fine no child SCCs, add them to the leaf list.
if (IsLeafSCC)
LeafSCCs.push_back(NewSCC);
return NewSCC;
}
LazyCallGraph::SCC *LazyCallGraph::getNextSCCInPostOrder() {
// When the stack is empty, there are no more SCCs to walk in this graph.
if (DFSStack.empty()) {
// If we've handled all candidate entry nodes to the SCC forest, we're done.
if (SCCEntryNodes.empty())
return nullptr;
// Reset the DFS numbering.
NextDFSNumber = 1;
Node &N = get(*SCCEntryNodes.pop_back_val());
DFSStack.push_back(std::make_pair(&N, N.begin()));
}
auto SI = DFSStack.rbegin();
if (SI->first->DFSNumber == 0) {
// This node hasn't been visited before, assign it a DFS number and remove
// it from the entry set.
assert(!SCCMap.count(SI->first) &&
[LCG] Add the first round of mutation support to the lazy call graph. This implements the core functionality necessary to remove an edge from the call graph and correctly update both the basic graph and the SCC structure. As part of that it has to run a tiny (in number of nodes) Tarjan-style DFS walk of an SCC being mutated to compute newly formed SCCs, etc. This is *very rough* and a WIP. I have a bunch of FIXMEs for code cleanup that will reduce the boilerplate in this change substantially. I also have a bunch of simplifications to various parts of both algorithms that I want to make, but first I'd like to have a more holistic picture. Ideally, I'd also like more testing. I'll probably add quite a few more unit tests as I go here to cover the various different aspects and corner cases of removing edges from the graph. Still, this is, so far, successfully updating the SCC graph in-place without disrupting the identity established for the existing SCCs even when we do challenging things like delete the critical edge that made an SCC cycle at all and have to reform things as a tree of smaller SCCs. Getting this to work is really critical for the new pass manager as it is going to associate significant state with the SCC instance and needs it to be stable. That is also the motivation behind the return of the newly formed SCCs. Eventually, I'll wire this all the way up to the public API so that the pass manager can use it to correctly re-enqueue newly formed SCCs into a fresh postorder traversal. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206968 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-23 11:03:03 +00:00
"Found a node with 0 DFS number but already in an SCC!");
SI->first->LowLink = SI->first->DFSNumber = NextDFSNumber++;
SCCEntryNodes.remove(&SI->first->getFunction());
}
do {
Node *N = SI->first;
for (auto I = SI->second, E = N->end(); I != E; ++I) {
Node *ChildN = *I;
if (ChildN->DFSNumber == 0) {
// Mark that we should start at this child when next this node is the
// top of the stack. We don't start at the next child to ensure this
// child's lowlink is reflected.
SI->second = I;
// Recurse onto this node via a tail call.
DFSStack.push_back(std::make_pair(ChildN, ChildN->begin()));
return LazyCallGraph::getNextSCCInPostOrder();
}
// Track the lowest link of the childen, if any are still in the stack.
assert(ChildN->LowLink != 0 &&
"Low-link must not be zero with a non-zero DFS number.");
if (ChildN->LowLink >= 0 && ChildN->LowLink < N->LowLink)
N->LowLink = ChildN->LowLink;
}
// No more children to process for this stack entry.
SI->second = N->end();
if (N->LowLink == N->DFSNumber)
// Form the new SCC out of the top of the DFS stack.
return formSCCFromDFSStack(DFSStack, std::prev(SI.base()));
++SI;
} while (SI != DFSStack.rend());
llvm_unreachable(
"We cannot reach the bottom of the stack without popping an SCC.");
}
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
char LazyCallGraphAnalysis::PassID;
LazyCallGraphPrinterPass::LazyCallGraphPrinterPass(raw_ostream &OS) : OS(OS) {}
static void printNodes(raw_ostream &OS, LazyCallGraph::Node &N,
SmallPtrSetImpl<LazyCallGraph::Node *> &Printed) {
// Recurse depth first through the nodes.
for (LazyCallGraph::Node *ChildN : N)
if (Printed.insert(ChildN))
printNodes(OS, *ChildN, Printed);
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
OS << " Call edges in function: " << N.getFunction().getName() << "\n";
for (LazyCallGraph::iterator I = N.begin(), E = N.end(); I != E; ++I)
OS << " -> " << I->getFunction().getName() << "\n";
OS << "\n";
}
[LCG] Add support for building persistent and connected SCCs to the LazyCallGraph. This is the start of the whole point of this different abstraction, but it is just the initial bits. Here is a run-down of what's going on here. I'm planning to incorporate some (or all) of this into comments going forward, hopefully with better editing and wording. =] The crux of the problem with the traditional way of building SCCs is that they are ephemeral. The new pass manager however really needs the ability to associate analysis passes and results of analysis passes with SCCs in order to expose these analysis passes to the SCC passes. Making this work is kind-of the whole point of the new pass manager. =] So, when we're building SCCs for the call graph, we actually want to build persistent nodes that stick around and can be reasoned about later. We'd also like the ability to walk the SCC graph in more complex ways than just the traditional postorder traversal of the current CGSCC walk. That means that in addition to being persistent, the SCCs need to be connected into a useful graph structure. However, we still want the SCCs to be formed lazily where possible. These constraints are quite hard to satisfy with the SCC iterator. Also, using that would bypass our ability to actually add data to the nodes of the call graph to facilite implementing the Tarjan walk. So I've re-implemented things in a more direct and embedded way. This immediately makes it easy to get the persistence and connectivity correct, and it also allows leveraging the existing nodes to simplify the algorithm. I've worked somewhat to make this implementation more closely follow the traditional paper's nomenclature and strategy, although it is still a bit obtuse because it isn't recursive, using an explicit stack and a tail call instead, and it is interruptable, resuming each time we need another SCC. The other tricky bit here, and what actually took almost all the time and trials and errors I spent building this, is exactly *what* graph structure to build for the SCCs. The naive thing to build is the call graph in its newly acyclic form. I wrote about 4 versions of this which did precisely this. Inevitably, when I experimented with them across various use cases, they became incredibly awkward. It was all implementable, but it felt like a complete wrong fit. Square peg, round hole. There were two overriding aspects that pushed me in a different direction: 1) We want to discover the SCC graph in a postorder fashion. That means the root node will be the *last* node we find. Using the call-SCC DAG as the graph structure of the SCCs results in an orphaned graph until we discover a root. 2) We will eventually want to walk the SCC graph in parallel, exploring distinct sub-graphs independently, and synchronizing at merge points. This again is not helped by the call-SCC DAG structure. The structure which, quite surprisingly, ended up being completely natural to use is the *inverse* of the call-SCC DAG. We add the leaf SCCs to the graph as "roots", and have edges to the caller SCCs. Once I switched to building this structure, everything just fell into place elegantly. Aside from general cleanups (there are FIXMEs and too few comments overall) that are still needed, the other missing piece of this is support for iterating across levels of the SCC graph. These will become useful for implementing #2, but they aren't an immediate priority. Once SCCs are in good shape, I'll be working on adding mutation support for incremental updates and adding the pass manager that this analysis enables. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206581 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 10:50:32 +00:00
static void printSCC(raw_ostream &OS, LazyCallGraph::SCC &SCC) {
ptrdiff_t SCCSize = std::distance(SCC.begin(), SCC.end());
OS << " SCC with " << SCCSize << " functions:\n";
for (LazyCallGraph::Node *N : SCC)
OS << " " << N->getFunction().getName() << "\n";
OS << "\n";
}
PreservedAnalyses LazyCallGraphPrinterPass::run(Module *M,
ModuleAnalysisManager *AM) {
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
LazyCallGraph &G = AM->getResult<LazyCallGraphAnalysis>(M);
OS << "Printing the call graph for module: " << M->getModuleIdentifier()
<< "\n\n";
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
SmallPtrSet<LazyCallGraph::Node *, 16> Printed;
for (LazyCallGraph::Node *N : G)
if (Printed.insert(N))
printNodes(OS, *N, Printed);
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
[LCG] Add support for building persistent and connected SCCs to the LazyCallGraph. This is the start of the whole point of this different abstraction, but it is just the initial bits. Here is a run-down of what's going on here. I'm planning to incorporate some (or all) of this into comments going forward, hopefully with better editing and wording. =] The crux of the problem with the traditional way of building SCCs is that they are ephemeral. The new pass manager however really needs the ability to associate analysis passes and results of analysis passes with SCCs in order to expose these analysis passes to the SCC passes. Making this work is kind-of the whole point of the new pass manager. =] So, when we're building SCCs for the call graph, we actually want to build persistent nodes that stick around and can be reasoned about later. We'd also like the ability to walk the SCC graph in more complex ways than just the traditional postorder traversal of the current CGSCC walk. That means that in addition to being persistent, the SCCs need to be connected into a useful graph structure. However, we still want the SCCs to be formed lazily where possible. These constraints are quite hard to satisfy with the SCC iterator. Also, using that would bypass our ability to actually add data to the nodes of the call graph to facilite implementing the Tarjan walk. So I've re-implemented things in a more direct and embedded way. This immediately makes it easy to get the persistence and connectivity correct, and it also allows leveraging the existing nodes to simplify the algorithm. I've worked somewhat to make this implementation more closely follow the traditional paper's nomenclature and strategy, although it is still a bit obtuse because it isn't recursive, using an explicit stack and a tail call instead, and it is interruptable, resuming each time we need another SCC. The other tricky bit here, and what actually took almost all the time and trials and errors I spent building this, is exactly *what* graph structure to build for the SCCs. The naive thing to build is the call graph in its newly acyclic form. I wrote about 4 versions of this which did precisely this. Inevitably, when I experimented with them across various use cases, they became incredibly awkward. It was all implementable, but it felt like a complete wrong fit. Square peg, round hole. There were two overriding aspects that pushed me in a different direction: 1) We want to discover the SCC graph in a postorder fashion. That means the root node will be the *last* node we find. Using the call-SCC DAG as the graph structure of the SCCs results in an orphaned graph until we discover a root. 2) We will eventually want to walk the SCC graph in parallel, exploring distinct sub-graphs independently, and synchronizing at merge points. This again is not helped by the call-SCC DAG structure. The structure which, quite surprisingly, ended up being completely natural to use is the *inverse* of the call-SCC DAG. We add the leaf SCCs to the graph as "roots", and have edges to the caller SCCs. Once I switched to building this structure, everything just fell into place elegantly. Aside from general cleanups (there are FIXMEs and too few comments overall) that are still needed, the other missing piece of this is support for iterating across levels of the SCC graph. These will become useful for implementing #2, but they aren't an immediate priority. Once SCCs are in good shape, I'll be working on adding mutation support for incremental updates and adding the pass manager that this analysis enables. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206581 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 10:50:32 +00:00
for (LazyCallGraph::SCC *SCC : G.postorder_sccs())
printSCC(OS, *SCC);
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
return PreservedAnalyses::all();
[LCG] Add support for building persistent and connected SCCs to the LazyCallGraph. This is the start of the whole point of this different abstraction, but it is just the initial bits. Here is a run-down of what's going on here. I'm planning to incorporate some (or all) of this into comments going forward, hopefully with better editing and wording. =] The crux of the problem with the traditional way of building SCCs is that they are ephemeral. The new pass manager however really needs the ability to associate analysis passes and results of analysis passes with SCCs in order to expose these analysis passes to the SCC passes. Making this work is kind-of the whole point of the new pass manager. =] So, when we're building SCCs for the call graph, we actually want to build persistent nodes that stick around and can be reasoned about later. We'd also like the ability to walk the SCC graph in more complex ways than just the traditional postorder traversal of the current CGSCC walk. That means that in addition to being persistent, the SCCs need to be connected into a useful graph structure. However, we still want the SCCs to be formed lazily where possible. These constraints are quite hard to satisfy with the SCC iterator. Also, using that would bypass our ability to actually add data to the nodes of the call graph to facilite implementing the Tarjan walk. So I've re-implemented things in a more direct and embedded way. This immediately makes it easy to get the persistence and connectivity correct, and it also allows leveraging the existing nodes to simplify the algorithm. I've worked somewhat to make this implementation more closely follow the traditional paper's nomenclature and strategy, although it is still a bit obtuse because it isn't recursive, using an explicit stack and a tail call instead, and it is interruptable, resuming each time we need another SCC. The other tricky bit here, and what actually took almost all the time and trials and errors I spent building this, is exactly *what* graph structure to build for the SCCs. The naive thing to build is the call graph in its newly acyclic form. I wrote about 4 versions of this which did precisely this. Inevitably, when I experimented with them across various use cases, they became incredibly awkward. It was all implementable, but it felt like a complete wrong fit. Square peg, round hole. There were two overriding aspects that pushed me in a different direction: 1) We want to discover the SCC graph in a postorder fashion. That means the root node will be the *last* node we find. Using the call-SCC DAG as the graph structure of the SCCs results in an orphaned graph until we discover a root. 2) We will eventually want to walk the SCC graph in parallel, exploring distinct sub-graphs independently, and synchronizing at merge points. This again is not helped by the call-SCC DAG structure. The structure which, quite surprisingly, ended up being completely natural to use is the *inverse* of the call-SCC DAG. We add the leaf SCCs to the graph as "roots", and have edges to the caller SCCs. Once I switched to building this structure, everything just fell into place elegantly. Aside from general cleanups (there are FIXMEs and too few comments overall) that are still needed, the other missing piece of this is support for iterating across levels of the SCC graph. These will become useful for implementing #2, but they aren't an immediate priority. Once SCCs are in good shape, I'll be working on adding mutation support for incremental updates and adding the pass manager that this analysis enables. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206581 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 10:50:32 +00:00
[PM] Add a new "lazy" call graph analysis pass for the new pass manager. The primary motivation for this pass is to separate the call graph analysis used by the new pass manager's CGSCC pass management from the existing call graph analysis pass. That analysis pass is (somewhat unfortunately) over-constrained by the existing CallGraphSCCPassManager requirements. Those requirements make it *really* hard to cleanly layer the needed functionality for the new pass manager on top of the existing analysis. However, there are also a bunch of things that the pass manager would specifically benefit from doing differently from the existing call graph analysis, and this new implementation tries to address several of them: - Be lazy about scanning function definitions. The existing pass eagerly scans the entire module to build the initial graph. This new pass is significantly more lazy, and I plan to push this even further to maximize locality during CGSCC walks. - Don't use a single synthetic node to partition functions with an indirect call from functions whose address is taken. This node creates a huge choke-point which would preclude good parallelization across the fanout of the SCC graph when we got to the point of looking at such changes to LLVM. - Use a memory dense and lightweight representation of the call graph rather than value handles and tracking call instructions. This will require explicit update calls instead of some updates working transparently, but should end up being significantly more efficient. The explicit update calls ended up being needed in many cases for the existing call graph so we don't really lose anything. - Doesn't explicitly model SCCs and thus doesn't provide an "identity" for an SCC which is stable across updates. This is essential for the new pass manager to work correctly. - Only form the graph necessary for traversing all of the functions in an SCC friendly order. This is a much simpler graph structure and should be more memory dense. It does limit the ways in which it is appropriate to use this analysis. I wish I had a better name than "call graph". I've commented extensively this aspect. This is still very much a WIP, in fact it is really just the initial bits. But it is about the fourth version of the initial bits that I've implemented with each of the others running into really frustrating problms. This looks like it will actually work and I'd like to split the actual complexity across commits for the sake of my reviewers. =] The rest of the implementation along with lots of wiring will follow somewhat more rapidly now that there is a good path forward. Naturally, this doesn't impact any of the existing optimizer. This code is specific to the new pass manager. A bunch of thanks are deserved for the various folks that have helped with the design of this, especially Nick Lewycky who actually sat with me to go through the fundamentals of the final version here. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200903 91177308-0d34-0410-b5e6-96231b3b80d8
2014-02-06 04:37:03 +00:00
}