Yi Kong | 878f994 | 2023-12-13 12:55:04 +0900 | [diff] [blame^] | 1 | //===- GCMetadata.h - Garbage collector metadata ----------------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file declares the GCFunctionInfo and GCModuleInfo classes, which are |
| 10 | // used as a communication channel from the target code generator to the target |
| 11 | // garbage collectors. This interface allows code generators and garbage |
| 12 | // collectors to be developed independently. |
| 13 | // |
| 14 | // The GCFunctionInfo class logs the data necessary to build a type accurate |
| 15 | // stack map. The code generator outputs: |
| 16 | // |
| 17 | // - Safe points as specified by the GCStrategy's NeededSafePoints. |
| 18 | // - Stack offsets for GC roots, as specified by calls to llvm.gcroot |
| 19 | // |
| 20 | // As a refinement, liveness analysis calculates the set of live roots at each |
| 21 | // safe point. Liveness analysis is not presently performed by the code |
| 22 | // generator, so all roots are assumed live. |
| 23 | // |
| 24 | // GCModuleInfo simply collects GCFunctionInfo instances for each Function as |
| 25 | // they are compiled. This accretion is necessary for collectors which must emit |
| 26 | // a stack map for the compilation unit as a whole. Therefore, GCFunctionInfo |
| 27 | // outlives the MachineFunction from which it is derived and must not refer to |
| 28 | // any code generator data structures. |
| 29 | // |
| 30 | //===----------------------------------------------------------------------===// |
| 31 | |
| 32 | #ifndef LLVM_CODEGEN_GCMETADATA_H |
| 33 | #define LLVM_CODEGEN_GCMETADATA_H |
| 34 | |
| 35 | #include "llvm/ADT/DenseMap.h" |
| 36 | #include "llvm/ADT/SmallVector.h" |
| 37 | #include "llvm/ADT/StringMap.h" |
| 38 | #include "llvm/ADT/StringRef.h" |
| 39 | #include "llvm/IR/DebugLoc.h" |
| 40 | #include "llvm/IR/GCStrategy.h" |
| 41 | #include "llvm/Pass.h" |
| 42 | #include <algorithm> |
| 43 | #include <cstddef> |
| 44 | #include <cstdint> |
| 45 | #include <memory> |
| 46 | #include <vector> |
| 47 | |
| 48 | namespace llvm { |
| 49 | |
| 50 | class Constant; |
| 51 | class Function; |
| 52 | class MCSymbol; |
| 53 | |
| 54 | /// GCPoint - Metadata for a collector-safe point in machine code. |
| 55 | /// |
| 56 | struct GCPoint { |
| 57 | MCSymbol *Label; ///< A label. |
| 58 | DebugLoc Loc; |
| 59 | |
| 60 | GCPoint(MCSymbol *L, DebugLoc DL) |
| 61 | : Label(L), Loc(std::move(DL)) {} |
| 62 | }; |
| 63 | |
| 64 | /// GCRoot - Metadata for a pointer to an object managed by the garbage |
| 65 | /// collector. |
| 66 | struct GCRoot { |
| 67 | int Num; ///< Usually a frame index. |
| 68 | int StackOffset = -1; ///< Offset from the stack pointer. |
| 69 | const Constant *Metadata; ///< Metadata straight from the call |
| 70 | ///< to llvm.gcroot. |
| 71 | |
| 72 | GCRoot(int N, const Constant *MD) : Num(N), Metadata(MD) {} |
| 73 | }; |
| 74 | |
| 75 | /// Garbage collection metadata for a single function. Currently, this |
| 76 | /// information only applies to GCStrategies which use GCRoot. |
| 77 | class GCFunctionInfo { |
| 78 | public: |
| 79 | using iterator = std::vector<GCPoint>::iterator; |
| 80 | using roots_iterator = std::vector<GCRoot>::iterator; |
| 81 | using live_iterator = std::vector<GCRoot>::const_iterator; |
| 82 | |
| 83 | private: |
| 84 | const Function &F; |
| 85 | GCStrategy &S; |
| 86 | uint64_t FrameSize; |
| 87 | std::vector<GCRoot> Roots; |
| 88 | std::vector<GCPoint> SafePoints; |
| 89 | |
| 90 | // FIXME: Liveness. A 2D BitVector, perhaps? |
| 91 | // |
| 92 | // BitVector Liveness; |
| 93 | // |
| 94 | // bool islive(int point, int root) = |
| 95 | // Liveness[point * SafePoints.size() + root] |
| 96 | // |
| 97 | // The bit vector is the more compact representation where >3.2% of roots |
| 98 | // are live per safe point (1.5% on 64-bit hosts). |
| 99 | |
| 100 | public: |
| 101 | GCFunctionInfo(const Function &F, GCStrategy &S); |
| 102 | ~GCFunctionInfo(); |
| 103 | |
| 104 | /// getFunction - Return the function to which this metadata applies. |
| 105 | const Function &getFunction() const { return F; } |
| 106 | |
| 107 | /// getStrategy - Return the GC strategy for the function. |
| 108 | GCStrategy &getStrategy() { return S; } |
| 109 | |
| 110 | /// addStackRoot - Registers a root that lives on the stack. Num is the |
| 111 | /// stack object ID for the alloca (if the code generator is |
| 112 | // using MachineFrameInfo). |
| 113 | void addStackRoot(int Num, const Constant *Metadata) { |
| 114 | Roots.push_back(GCRoot(Num, Metadata)); |
| 115 | } |
| 116 | |
| 117 | /// removeStackRoot - Removes a root. |
| 118 | roots_iterator removeStackRoot(roots_iterator position) { |
| 119 | return Roots.erase(position); |
| 120 | } |
| 121 | |
| 122 | /// addSafePoint - Notes the existence of a safe point. Num is the ID of the |
| 123 | /// label just prior to the safe point (if the code generator is using |
| 124 | /// MachineModuleInfo). |
| 125 | void addSafePoint(MCSymbol *Label, const DebugLoc &DL) { |
| 126 | SafePoints.emplace_back(Label, DL); |
| 127 | } |
| 128 | |
| 129 | /// getFrameSize/setFrameSize - Records the function's frame size. |
| 130 | uint64_t getFrameSize() const { return FrameSize; } |
| 131 | void setFrameSize(uint64_t S) { FrameSize = S; } |
| 132 | |
| 133 | /// begin/end - Iterators for safe points. |
| 134 | iterator begin() { return SafePoints.begin(); } |
| 135 | iterator end() { return SafePoints.end(); } |
| 136 | size_t size() const { return SafePoints.size(); } |
| 137 | |
| 138 | /// roots_begin/roots_end - Iterators for all roots in the function. |
| 139 | roots_iterator roots_begin() { return Roots.begin(); } |
| 140 | roots_iterator roots_end() { return Roots.end(); } |
| 141 | size_t roots_size() const { return Roots.size(); } |
| 142 | |
| 143 | /// live_begin/live_end - Iterators for live roots at a given safe point. |
| 144 | live_iterator live_begin(const iterator &p) { return roots_begin(); } |
| 145 | live_iterator live_end(const iterator &p) { return roots_end(); } |
| 146 | size_t live_size(const iterator &p) const { return roots_size(); } |
| 147 | }; |
| 148 | |
| 149 | /// An analysis pass which caches information about the entire Module. |
| 150 | /// Records both the function level information used by GCRoots and a |
| 151 | /// cache of the 'active' gc strategy objects for the current Module. |
| 152 | class GCModuleInfo : public ImmutablePass { |
| 153 | /// An owning list of all GCStrategies which have been created |
| 154 | SmallVector<std::unique_ptr<GCStrategy>, 1> GCStrategyList; |
| 155 | /// A helper map to speedup lookups into the above list |
| 156 | StringMap<GCStrategy*> GCStrategyMap; |
| 157 | |
| 158 | public: |
| 159 | /// Lookup the GCStrategy object associated with the given gc name. |
| 160 | /// Objects are owned internally; No caller should attempt to delete the |
| 161 | /// returned objects. |
| 162 | GCStrategy *getGCStrategy(const StringRef Name); |
| 163 | |
| 164 | /// List of per function info objects. In theory, Each of these |
| 165 | /// may be associated with a different GC. |
| 166 | using FuncInfoVec = std::vector<std::unique_ptr<GCFunctionInfo>>; |
| 167 | |
| 168 | FuncInfoVec::iterator funcinfo_begin() { return Functions.begin(); } |
| 169 | FuncInfoVec::iterator funcinfo_end() { return Functions.end(); } |
| 170 | |
| 171 | private: |
| 172 | /// Owning list of all GCFunctionInfos associated with this Module |
| 173 | FuncInfoVec Functions; |
| 174 | |
| 175 | /// Non-owning map to bypass linear search when finding the GCFunctionInfo |
| 176 | /// associated with a particular Function. |
| 177 | using finfo_map_type = DenseMap<const Function *, GCFunctionInfo *>; |
| 178 | finfo_map_type FInfoMap; |
| 179 | |
| 180 | public: |
| 181 | using iterator = SmallVector<std::unique_ptr<GCStrategy>, 1>::const_iterator; |
| 182 | |
| 183 | static char ID; |
| 184 | |
| 185 | GCModuleInfo(); |
| 186 | |
| 187 | /// clear - Resets the pass. Any pass, which uses GCModuleInfo, should |
| 188 | /// call it in doFinalization(). |
| 189 | /// |
| 190 | void clear(); |
| 191 | |
| 192 | /// begin/end - Iterators for used strategies. |
| 193 | /// |
| 194 | iterator begin() const { return GCStrategyList.begin(); } |
| 195 | iterator end() const { return GCStrategyList.end(); } |
| 196 | |
| 197 | /// get - Look up function metadata. This is currently assumed |
| 198 | /// have the side effect of initializing the associated GCStrategy. That |
| 199 | /// will soon change. |
| 200 | GCFunctionInfo &getFunctionInfo(const Function &F); |
| 201 | }; |
| 202 | |
| 203 | } // end namespace llvm |
| 204 | |
| 205 | #endif // LLVM_CODEGEN_GCMETADATA_H |