GCMetadata.h revision 321369
1//===- GCMetadata.h - Garbage collector metadata ----------------*- C++ -*-===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file declares the GCFunctionInfo and GCModuleInfo classes, which are
11// used as a communication channel from the target code generator to the target
12// garbage collectors. This interface allows code generators and garbage
13// collectors to be developed independently.
14//
15// The GCFunctionInfo class logs the data necessary to build a type accurate
16// stack map. The code generator outputs:
17//
18//   - Safe points as specified by the GCStrategy's NeededSafePoints.
19//   - Stack offsets for GC roots, as specified by calls to llvm.gcroot
20//
21// As a refinement, liveness analysis calculates the set of live roots at each
22// safe point. Liveness analysis is not presently performed by the code
23// generator, so all roots are assumed live.
24//
25// GCModuleInfo simply collects GCFunctionInfo instances for each Function as
26// they are compiled. This accretion is necessary for collectors which must emit
27// a stack map for the compilation unit as a whole. Therefore, GCFunctionInfo
28// outlives the MachineFunction from which it is derived and must not refer to
29// any code generator data structures.
30//
31//===----------------------------------------------------------------------===//
32
33#ifndef LLVM_CODEGEN_GCMETADATA_H
34#define LLVM_CODEGEN_GCMETADATA_H
35
36#include "llvm/ADT/DenseMap.h"
37#include "llvm/ADT/SmallVector.h"
38#include "llvm/ADT/StringMap.h"
39#include "llvm/ADT/StringRef.h"
40#include "llvm/CodeGen/GCStrategy.h"
41#include "llvm/IR/DebugLoc.h"
42#include "llvm/Pass.h"
43#include <algorithm>
44#include <cstddef>
45#include <cstdint>
46#include <memory>
47#include <vector>
48
49namespace llvm {
50
51class Constant;
52class Function;
53class MCSymbol;
54
55/// GCPoint - Metadata for a collector-safe point in machine code.
56///
57struct GCPoint {
58  GC::PointKind Kind; ///< The kind of the safe point.
59  MCSymbol *Label;    ///< A label.
60  DebugLoc Loc;
61
62  GCPoint(GC::PointKind K, MCSymbol *L, DebugLoc DL)
63      : Kind(K), Label(L), Loc(std::move(DL)) {}
64};
65
66/// GCRoot - Metadata for a pointer to an object managed by the garbage
67/// collector.
68struct GCRoot {
69  int Num;                  ///< Usually a frame index.
70  int StackOffset = -1;     ///< Offset from the stack pointer.
71  const Constant *Metadata; ///< Metadata straight from the call
72                            ///< to llvm.gcroot.
73
74  GCRoot(int N, const Constant *MD) : Num(N), Metadata(MD) {}
75};
76
77/// Garbage collection metadata for a single function.  Currently, this
78/// information only applies to GCStrategies which use GCRoot.
79class GCFunctionInfo {
80public:
81  using iterator = std::vector<GCPoint>::iterator;
82  using roots_iterator = std::vector<GCRoot>::iterator;
83  using live_iterator = std::vector<GCRoot>::const_iterator;
84
85private:
86  const Function &F;
87  GCStrategy &S;
88  uint64_t FrameSize;
89  std::vector<GCRoot> Roots;
90  std::vector<GCPoint> SafePoints;
91
92  // FIXME: Liveness. A 2D BitVector, perhaps?
93  //
94  //   BitVector Liveness;
95  //
96  //   bool islive(int point, int root) =
97  //     Liveness[point * SafePoints.size() + root]
98  //
99  // The bit vector is the more compact representation where >3.2% of roots
100  // are live per safe point (1.5% on 64-bit hosts).
101
102public:
103  GCFunctionInfo(const Function &F, GCStrategy &S);
104  ~GCFunctionInfo();
105
106  /// getFunction - Return the function to which this metadata applies.
107  const Function &getFunction() const { return F; }
108
109  /// getStrategy - Return the GC strategy for the function.
110  GCStrategy &getStrategy() { return S; }
111
112  /// addStackRoot - Registers a root that lives on the stack. Num is the
113  ///                stack object ID for the alloca (if the code generator is
114  //                 using  MachineFrameInfo).
115  void addStackRoot(int Num, const Constant *Metadata) {
116    Roots.push_back(GCRoot(Num, Metadata));
117  }
118
119  /// removeStackRoot - Removes a root.
120  roots_iterator removeStackRoot(roots_iterator position) {
121    return Roots.erase(position);
122  }
123
124  /// addSafePoint - Notes the existence of a safe point. Num is the ID of the
125  /// label just prior to the safe point (if the code generator is using
126  /// MachineModuleInfo).
127  void addSafePoint(GC::PointKind Kind, MCSymbol *Label, const DebugLoc &DL) {
128    SafePoints.emplace_back(Kind, Label, DL);
129  }
130
131  /// getFrameSize/setFrameSize - Records the function's frame size.
132  uint64_t getFrameSize() const { return FrameSize; }
133  void setFrameSize(uint64_t S) { FrameSize = S; }
134
135  /// begin/end - Iterators for safe points.
136  iterator begin() { return SafePoints.begin(); }
137  iterator end() { return SafePoints.end(); }
138  size_t size() const { return SafePoints.size(); }
139
140  /// roots_begin/roots_end - Iterators for all roots in the function.
141  roots_iterator roots_begin() { return Roots.begin(); }
142  roots_iterator roots_end() { return Roots.end(); }
143  size_t roots_size() const { return Roots.size(); }
144
145  /// live_begin/live_end - Iterators for live roots at a given safe point.
146  live_iterator live_begin(const iterator &p) { return roots_begin(); }
147  live_iterator live_end(const iterator &p) { return roots_end(); }
148  size_t live_size(const iterator &p) const { return roots_size(); }
149};
150
151/// An analysis pass which caches information about the entire Module.
152/// Records both the function level information used by GCRoots and a
153/// cache of the 'active' gc strategy objects for the current Module.
154class GCModuleInfo : public ImmutablePass {
155  /// An owning list of all GCStrategies which have been created
156  SmallVector<std::unique_ptr<GCStrategy>, 1> GCStrategyList;
157  /// A helper map to speedup lookups into the above list
158  StringMap<GCStrategy*> GCStrategyMap;
159
160public:
161  /// Lookup the GCStrategy object associated with the given gc name.
162  /// Objects are owned internally; No caller should attempt to delete the
163  /// returned objects.
164  GCStrategy *getGCStrategy(const StringRef Name);
165
166  /// List of per function info objects.  In theory, Each of these
167  /// may be associated with a different GC.
168  using FuncInfoVec = std::vector<std::unique_ptr<GCFunctionInfo>>;
169
170  FuncInfoVec::iterator funcinfo_begin() { return Functions.begin(); }
171  FuncInfoVec::iterator funcinfo_end() { return Functions.end(); }
172
173private:
174  /// Owning list of all GCFunctionInfos associated with this Module
175  FuncInfoVec Functions;
176
177  /// Non-owning map to bypass linear search when finding the GCFunctionInfo
178  /// associated with a particular Function.
179  using finfo_map_type = DenseMap<const Function *, GCFunctionInfo *>;
180  finfo_map_type FInfoMap;
181
182public:
183  using iterator = SmallVector<std::unique_ptr<GCStrategy>, 1>::const_iterator;
184
185  static char ID;
186
187  GCModuleInfo();
188
189  /// clear - Resets the pass. Any pass, which uses GCModuleInfo, should
190  /// call it in doFinalization().
191  ///
192  void clear();
193
194  /// begin/end - Iterators for used strategies.
195  ///
196  iterator begin() const { return GCStrategyList.begin(); }
197  iterator end() const { return GCStrategyList.end(); }
198
199  /// get - Look up function metadata.  This is currently assumed
200  /// have the side effect of initializing the associated GCStrategy.  That
201  /// will soon change.
202  GCFunctionInfo &getFunctionInfo(const Function &F);
203};
204
205} // end namespace llvm
206
207#endif // LLVM_CODEGEN_GCMETADATA_H
208