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/CodeGen/GCStrategy.h"
40 #include "llvm/IR/DebugLoc.h"
41 #include "llvm/Pass.h"
42 #include <memory>
43 
44 namespace llvm {
45 class AsmPrinter;
46 class Constant;
47 class MCSymbol;
48 
49 /// GCPoint - Metadata for a collector-safe point in machine code.
50 ///
51 struct GCPoint {
52   GC::PointKind Kind; ///< The kind of the safe point.
53   MCSymbol *Label;    ///< A label.
54   DebugLoc Loc;
55 
GCPointGCPoint56   GCPoint(GC::PointKind K, MCSymbol *L, DebugLoc DL)
57       : Kind(K), Label(L), Loc(DL) {}
58 };
59 
60 /// GCRoot - Metadata for a pointer to an object managed by the garbage
61 /// collector.
62 struct GCRoot {
63   int Num;                  ///< Usually a frame index.
64   int StackOffset;          ///< Offset from the stack pointer.
65   const Constant *Metadata; ///< Metadata straight from the call
66                             ///< to llvm.gcroot.
67 
GCRootGCRoot68   GCRoot(int N, const Constant *MD) : Num(N), StackOffset(-1), Metadata(MD) {}
69 };
70 
71 /// Garbage collection metadata for a single function.  Currently, this
72 /// information only applies to GCStrategies which use GCRoot.
73 class GCFunctionInfo {
74 public:
75   typedef std::vector<GCPoint>::iterator iterator;
76   typedef std::vector<GCRoot>::iterator roots_iterator;
77   typedef std::vector<GCRoot>::const_iterator live_iterator;
78 
79 private:
80   const Function &F;
81   GCStrategy &S;
82   uint64_t FrameSize;
83   std::vector<GCRoot> Roots;
84   std::vector<GCPoint> SafePoints;
85 
86   // FIXME: Liveness. A 2D BitVector, perhaps?
87   //
88   //   BitVector Liveness;
89   //
90   //   bool islive(int point, int root) =
91   //     Liveness[point * SafePoints.size() + root]
92   //
93   // The bit vector is the more compact representation where >3.2% of roots
94   // are live per safe point (1.5% on 64-bit hosts).
95 
96 public:
97   GCFunctionInfo(const Function &F, GCStrategy &S);
98   ~GCFunctionInfo();
99 
100   /// getFunction - Return the function to which this metadata applies.
101   ///
getFunction()102   const Function &getFunction() const { return F; }
103 
104   /// getStrategy - Return the GC strategy for the function.
105   ///
getStrategy()106   GCStrategy &getStrategy() { return S; }
107 
108   /// addStackRoot - Registers a root that lives on the stack. Num is the
109   ///                stack object ID for the alloca (if the code generator is
110   //                 using  MachineFrameInfo).
addStackRoot(int Num,const Constant * Metadata)111   void addStackRoot(int Num, const Constant *Metadata) {
112     Roots.push_back(GCRoot(Num, Metadata));
113   }
114 
115   /// removeStackRoot - Removes a root.
removeStackRoot(roots_iterator position)116   roots_iterator removeStackRoot(roots_iterator position) {
117     return Roots.erase(position);
118   }
119 
120   /// addSafePoint - Notes the existence of a safe point. Num is the ID of the
121   /// label just prior to the safe point (if the code generator is using
122   /// MachineModuleInfo).
addSafePoint(GC::PointKind Kind,MCSymbol * Label,DebugLoc DL)123   void addSafePoint(GC::PointKind Kind, MCSymbol *Label, DebugLoc DL) {
124     SafePoints.emplace_back(Kind, Label, DL);
125   }
126 
127   /// getFrameSize/setFrameSize - Records the function's frame size.
128   ///
getFrameSize()129   uint64_t getFrameSize() const { return FrameSize; }
setFrameSize(uint64_t S)130   void setFrameSize(uint64_t S) { FrameSize = S; }
131 
132   /// begin/end - Iterators for safe points.
133   ///
begin()134   iterator begin() { return SafePoints.begin(); }
end()135   iterator end() { return SafePoints.end(); }
size()136   size_t size() const { return SafePoints.size(); }
137 
138   /// roots_begin/roots_end - Iterators for all roots in the function.
139   ///
roots_begin()140   roots_iterator roots_begin() { return Roots.begin(); }
roots_end()141   roots_iterator roots_end() { return Roots.end(); }
roots_size()142   size_t roots_size() const { return Roots.size(); }
143 
144   /// live_begin/live_end - Iterators for live roots at a given safe point.
145   ///
live_begin(const iterator & p)146   live_iterator live_begin(const iterator &p) { return roots_begin(); }
live_end(const iterator & p)147   live_iterator live_end(const iterator &p) { return roots_end(); }
live_size(const iterator & p)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.
154 class 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 
160 public:
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   typedef std::vector<std::unique_ptr<GCFunctionInfo>> FuncInfoVec;
169 
funcinfo_begin()170   FuncInfoVec::iterator funcinfo_begin() { return Functions.begin(); }
funcinfo_end()171   FuncInfoVec::iterator funcinfo_end() { return Functions.end(); }
172 
173 private:
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   typedef DenseMap<const Function *, GCFunctionInfo *> finfo_map_type;
180   finfo_map_type FInfoMap;
181 
182 public:
183   typedef SmallVector<std::unique_ptr<GCStrategy>,1>::const_iterator 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   ///
begin()196   iterator begin() const { return GCStrategyList.begin(); }
end()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 
206 #endif
207