1 //===- LazyCallGraph.h - Analysis of a Module's call graph ------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// Implements a lazy call graph analysis and related passes for the new pass
14 /// NB: This is *not* a traditional call graph! It is a graph which models both
15 /// the current calls and potential calls. As a consequence there are many
16 /// edges in this call graph that do not correspond to a 'call' or 'invoke'
19 /// The primary use cases of this graph analysis is to facilitate iterating
20 /// across the functions of a module in ways that ensure all callees are
21 /// visited prior to a caller (given any SCC constraints), or vice versa. As
22 /// such is it particularly well suited to organizing CGSCC optimizations such
23 /// as inlining, outlining, argument promotion, etc. That is its primary use
24 /// case and motivates the design. It may not be appropriate for other
25 /// purposes. The use graph of functions or some other conservative analysis of
26 /// call instructions may be interesting for optimizations and subsequent
27 /// analyses which don't work in the context of an overly specified
28 /// potential-call-edge graph.
30 /// To understand the specific rules and nature of this call graph analysis,
31 /// see the documentation of the \c LazyCallGraph below.
33 //===----------------------------------------------------------------------===//
35 #ifndef LLVM_ANALYSIS_LAZY_CALL_GRAPH
36 #define LLVM_ANALYSIS_LAZY_CALL_GRAPH
38 #include "llvm/ADT/DenseMap.h"
39 #include "llvm/ADT/PointerUnion.h"
40 #include "llvm/ADT/STLExtras.h"
41 #include "llvm/ADT/SetVector.h"
42 #include "llvm/ADT/SmallPtrSet.h"
43 #include "llvm/ADT/SmallVector.h"
44 #include "llvm/ADT/iterator_range.h"
45 #include "llvm/IR/BasicBlock.h"
46 #include "llvm/IR/Function.h"
47 #include "llvm/IR/Module.h"
48 #include "llvm/Support/Allocator.h"
52 class ModuleAnalysisManager;
53 class PreservedAnalyses;
56 /// \brief A lazily constructed view of the call graph of a module.
58 /// With the edges of this graph, the motivating constraint that we are
59 /// attempting to maintain is that function-local optimization, CGSCC-local
60 /// optimizations, and optimizations transforming a pair of functions connected
61 /// by an edge in the graph, do not invalidate a bottom-up traversal of the SCC
62 /// DAG. That is, no optimizations will delete, remove, or add an edge such
63 /// that functions already visited in a bottom-up order of the SCC DAG are no
64 /// longer valid to have visited, or such that functions not yet visited in
65 /// a bottom-up order of the SCC DAG are not required to have already been
68 /// Within this constraint, the desire is to minimize the merge points of the
69 /// SCC DAG. The greater the fanout of the SCC DAG and the fewer merge points
70 /// in the SCC DAG, the more independence there is in optimizing within it.
71 /// There is a strong desire to enable parallelization of optimizations over
72 /// the call graph, and both limited fanout and merge points will (artificially
73 /// in some cases) limit the scaling of such an effort.
75 /// To this end, graph represents both direct and any potential resolution to
76 /// an indirect call edge. Another way to think about it is that it represents
77 /// both the direct call edges and any direct call edges that might be formed
78 /// through static optimizations. Specifically, it considers taking the address
79 /// of a function to be an edge in the call graph because this might be
80 /// forwarded to become a direct call by some subsequent function-local
81 /// optimization. The result is that the graph closely follows the use-def
82 /// edges for functions. Walking "up" the graph can be done by looking at all
83 /// of the uses of a function.
85 /// The roots of the call graph are the external functions and functions
86 /// escaped into global variables. Those functions can be called from outside
87 /// of the module or via unknowable means in the IR -- we may not be able to
88 /// form even a potential call edge from a function body which may dynamically
89 /// load the function and call it.
91 /// This analysis still requires updates to remain valid after optimizations
92 /// which could potentially change the set of potential callees. The
93 /// constraints it operates under only make the traversal order remain valid.
95 /// The entire analysis must be re-computed if full interprocedural
96 /// optimizations run at any point. For example, globalopt completely
97 /// invalidates the information in this analysis.
99 /// FIXME: This class is named LazyCallGraph in a lame attempt to distinguish
100 /// it from the existing CallGraph. At some point, it is expected that this
101 /// will be the only call graph and it will be renamed accordingly.
102 class LazyCallGraph {
106 typedef SmallVector<PointerUnion<Function *, Node *>, 4> NodeVectorT;
107 typedef SmallVectorImpl<PointerUnion<Function *, Node *>> NodeVectorImplT;
109 /// \brief A lazy iterator used for both the entry nodes and child nodes.
111 /// When this iterator is dereferenced, if not yet available, a function will
112 /// be scanned for "calls" or uses of functions and its child information
113 /// will be constructed. All of these results are accumulated and cached in
115 class iterator : public std::iterator<std::bidirectional_iterator_tag, Node *,
116 ptrdiff_t, Node *, Node *> {
117 friend class LazyCallGraph;
118 friend class LazyCallGraph::Node;
119 typedef std::iterator<std::bidirectional_iterator_tag, Node *, ptrdiff_t,
120 Node *, Node *> BaseT;
122 /// \brief Nonce type to select the constructor for the end iterator.
126 NodeVectorImplT::iterator NI;
128 // Build the begin iterator for a node.
129 explicit iterator(LazyCallGraph &G, NodeVectorImplT &Nodes)
130 : G(&G), NI(Nodes.begin()) {}
132 // Build the end iterator for a node. This is selected purely by overload.
133 iterator(LazyCallGraph &G, NodeVectorImplT &Nodes, IsAtEndT /*Nonce*/)
134 : G(&G), NI(Nodes.end()) {}
137 bool operator==(const iterator &Arg) { return NI == Arg.NI; }
138 bool operator!=(const iterator &Arg) { return !operator==(Arg); }
140 reference operator*() const {
141 if (NI->is<Node *>())
142 return NI->get<Node *>();
144 Function *F = NI->get<Function *>();
145 Node *ChildN = G->get(*F);
149 pointer operator->() const { return operator*(); }
151 iterator &operator++() {
155 iterator operator++(int) {
156 iterator prev = *this;
161 iterator &operator--() {
165 iterator operator--(int) {
166 iterator next = *this;
172 /// \brief A node in the call graph.
174 /// This represents a single node. It's primary roles are to cache the list of
175 /// callees, de-duplicate and provide fast testing of whether a function is
176 /// a callee, and facilitate iteration of child nodes in the graph.
178 friend class LazyCallGraph;
179 friend class LazyCallGraph::SCC;
184 // We provide for the DFS numbering and Tarjan walk lowlink numbers to be
185 // stored directly within the node.
189 mutable NodeVectorT Callees;
190 SmallPtrSet<Function *, 4> CalleeSet;
192 /// \brief Basic constructor implements the scanning of F into Callees and
194 Node(LazyCallGraph &G, Function &F);
196 /// \brief Constructor used when copying a node from one graph to another.
197 Node(LazyCallGraph &G, const Node &OtherN);
200 typedef LazyCallGraph::iterator iterator;
202 Function &getFunction() const {
206 iterator begin() const { return iterator(*G, Callees); }
207 iterator end() const { return iterator(*G, Callees, iterator::IsAtEndT()); }
209 /// Equality is defined as address equality.
210 bool operator==(const Node &N) const { return this == &N; }
211 bool operator!=(const Node &N) const { return !operator==(N); }
214 /// \brief An SCC of the call graph.
216 /// This represents a Strongly Connected Component of the call graph as
217 /// a collection of call graph nodes. While the order of nodes in the SCC is
218 /// stable, it is not any particular order.
220 friend class LazyCallGraph;
221 friend class LazyCallGraph::Node;
223 SmallSetVector<SCC *, 1> ParentSCCs;
224 SmallVector<Node *, 1> Nodes;
225 SmallPtrSet<Function *, 1> NodeSet;
230 typedef SmallVectorImpl<Node *>::const_iterator iterator;
232 iterator begin() const { return Nodes.begin(); }
233 iterator end() const { return Nodes.end(); }
236 /// \brief A post-order depth-first SCC iterator over the call graph.
238 /// This iterator triggers the Tarjan DFS-based formation of the SCC DAG for
239 /// the call graph, walking it lazily in depth-first post-order. That is, it
240 /// always visits SCCs for a callee prior to visiting the SCC for a caller
241 /// (when they are in different SCCs).
242 class postorder_scc_iterator
243 : public std::iterator<std::forward_iterator_tag, SCC *, ptrdiff_t, SCC *,
245 friend class LazyCallGraph;
246 friend class LazyCallGraph::Node;
247 typedef std::iterator<std::forward_iterator_tag, SCC *, ptrdiff_t,
250 /// \brief Nonce type to select the constructor for the end iterator.
256 // Build the begin iterator for a node.
257 postorder_scc_iterator(LazyCallGraph &G) : G(&G) {
258 C = G.getNextSCCInPostOrder();
261 // Build the end iterator for a node. This is selected purely by overload.
262 postorder_scc_iterator(LazyCallGraph &G, IsAtEndT /*Nonce*/)
263 : G(&G), C(nullptr) {}
266 bool operator==(const postorder_scc_iterator &Arg) {
267 return G == Arg.G && C == Arg.C;
269 bool operator!=(const postorder_scc_iterator &Arg) {
270 return !operator==(Arg);
273 reference operator*() const { return C; }
274 pointer operator->() const { return operator*(); }
276 postorder_scc_iterator &operator++() {
277 C = G->getNextSCCInPostOrder();
280 postorder_scc_iterator operator++(int) {
281 postorder_scc_iterator prev = *this;
287 /// \brief Construct a graph for the given module.
289 /// This sets up the graph and computes all of the entry points of the graph.
290 /// No function definitions are scanned until their nodes in the graph are
291 /// requested during traversal.
292 LazyCallGraph(Module &M);
294 /// \brief Copy constructor.
296 /// This does a deep copy of the graph. It does no verification that the
297 /// graph remains valid for the module. It is also relatively expensive.
298 LazyCallGraph(const LazyCallGraph &G);
300 /// \brief Move constructor.
302 /// This is a deep move. It leaves G in an undefined but destroyable state.
303 /// Any other operation on G is likely to fail.
304 LazyCallGraph(LazyCallGraph &&G);
306 /// \brief Copy and move assignment.
307 LazyCallGraph &operator=(LazyCallGraph RHS) {
308 std::swap(*this, RHS);
312 iterator begin() { return iterator(*this, EntryNodes); }
313 iterator end() { return iterator(*this, EntryNodes, iterator::IsAtEndT()); }
315 postorder_scc_iterator postorder_scc_begin() {
316 return postorder_scc_iterator(*this);
318 postorder_scc_iterator postorder_scc_end() {
319 return postorder_scc_iterator(*this, postorder_scc_iterator::IsAtEndT());
322 iterator_range<postorder_scc_iterator> postorder_sccs() {
323 return iterator_range<postorder_scc_iterator>(postorder_scc_begin(),
324 postorder_scc_end());
327 /// \brief Lookup a function in the graph which has already been scanned and
329 Node *lookup(const Function &F) const { return NodeMap.lookup(&F); }
331 /// \brief Get a graph node for a given function, scanning it to populate the
332 /// graph data as necessary.
333 Node *get(Function &F) {
334 Node *&N = NodeMap[&F];
338 return insertInto(F, N);
342 /// \brief Allocator that holds all the call graph nodes.
343 SpecificBumpPtrAllocator<Node> BPA;
345 /// \brief Maps function->node for fast lookup.
346 DenseMap<const Function *, Node *> NodeMap;
348 /// \brief The entry nodes to the graph.
350 /// These nodes are reachable through "external" means. Put another way, they
351 /// escape at the module scope.
352 NodeVectorT EntryNodes;
354 /// \brief Set of the entry nodes to the graph.
355 SmallPtrSet<Function *, 4> EntryNodeSet;
357 /// \brief Allocator that holds all the call graph SCCs.
358 SpecificBumpPtrAllocator<SCC> SCCBPA;
360 /// \brief Maps Function -> SCC for fast lookup.
361 DenseMap<const Function *, SCC *> SCCMap;
363 /// \brief The leaf SCCs of the graph.
365 /// These are all of the SCCs which have no children.
366 SmallVector<SCC *, 4> LeafSCCs;
368 /// \brief Stack of nodes not-yet-processed into SCCs.
369 SmallVector<std::pair<Node *, iterator>, 4> DFSStack;
371 /// \brief Set of entry nodes not-yet-processed into SCCs.
372 SmallSetVector<Function *, 4> SCCEntryNodes;
374 /// \brief Counter for the next DFS number to assign.
377 /// \brief Helper to insert a new function, with an already looked-up entry in
379 Node *insertInto(Function &F, Node *&MappedN);
381 /// \brief Helper to copy a node from another graph into this one.
382 Node *copyInto(const Node &OtherN);
384 /// \brief Retrieve the next node in the post-order SCC walk of the call graph.
385 SCC *getNextSCCInPostOrder();
388 // Provide GraphTraits specializations for call graphs.
389 template <> struct GraphTraits<LazyCallGraph::Node *> {
390 typedef LazyCallGraph::Node NodeType;
391 typedef LazyCallGraph::iterator ChildIteratorType;
393 static NodeType *getEntryNode(NodeType *N) { return N; }
394 static ChildIteratorType child_begin(NodeType *N) { return N->begin(); }
395 static ChildIteratorType child_end(NodeType *N) { return N->end(); }
397 template <> struct GraphTraits<LazyCallGraph *> {
398 typedef LazyCallGraph::Node NodeType;
399 typedef LazyCallGraph::iterator ChildIteratorType;
401 static NodeType *getEntryNode(NodeType *N) { return N; }
402 static ChildIteratorType child_begin(NodeType *N) { return N->begin(); }
403 static ChildIteratorType child_end(NodeType *N) { return N->end(); }
406 /// \brief An analysis pass which computes the call graph for a module.
407 class LazyCallGraphAnalysis {
409 /// \brief Inform generic clients of the result type.
410 typedef LazyCallGraph Result;
412 static void *ID() { return (void *)&PassID; }
414 /// \brief Compute the \c LazyCallGraph for a the module \c M.
416 /// This just builds the set of entry points to the call graph. The rest is
417 /// built lazily as it is walked.
418 LazyCallGraph run(Module *M) { return LazyCallGraph(*M); }
424 /// \brief A pass which prints the call graph to a \c raw_ostream.
426 /// This is primarily useful for testing the analysis.
427 class LazyCallGraphPrinterPass {
431 explicit LazyCallGraphPrinterPass(raw_ostream &OS);
433 PreservedAnalyses run(Module *M, ModuleAnalysisManager *AM);
435 static StringRef name() { return "LazyCallGraphPrinterPass"; }