Home | History | Annotate | Download | only in MCJIT 
      1 //===-- MCJIT.h - Class definition for the MCJIT ----------------*- 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 #ifndef LLVM_LIB_EXECUTIONENGINE_MCJIT_H
     11 #define LLVM_LIB_EXECUTIONENGINE_MCJIT_H
     12 
     13 #include "llvm/ADT/DenseMap.h"
     14 #include "llvm/ADT/SmallPtrSet.h"
     15 #include "llvm/ADT/SmallVector.h"
     16 #include "llvm/ExecutionEngine/ExecutionEngine.h"
     17 #include "llvm/ExecutionEngine/ObjectCache.h"
     18 #include "llvm/ExecutionEngine/ObjectImage.h"
     19 #include "llvm/ExecutionEngine/RuntimeDyld.h"
     20 #include "llvm/IR/Module.h"
     21 
     22 namespace llvm {
     23 class MCJIT;
     24 
     25 // This is a helper class that the MCJIT execution engine uses for linking
     26 // functions across modules that it owns.  It aggregates the memory manager
     27 // that is passed in to the MCJIT constructor and defers most functionality
     28 // to that object.
     29 class LinkingMemoryManager : public RTDyldMemoryManager {
     30 public:
     31   LinkingMemoryManager(MCJIT *Parent, RTDyldMemoryManager *MM)
     32     : ParentEngine(Parent), ClientMM(MM) {}
     33 
     34   uint64_t getSymbolAddress(const std::string &Name) override;
     35 
     36   // Functions deferred to client memory manager
     37   uint8_t *allocateCodeSection(uintptr_t Size, unsigned Alignment,
     38                                unsigned SectionID,
     39                                StringRef SectionName) override {
     40     return ClientMM->allocateCodeSection(Size, Alignment, SectionID, SectionName);
     41   }
     42 
     43   uint8_t *allocateDataSection(uintptr_t Size, unsigned Alignment,
     44                                unsigned SectionID, StringRef SectionName,
     45                                bool IsReadOnly) override {
     46     return ClientMM->allocateDataSection(Size, Alignment,
     47                                          SectionID, SectionName, IsReadOnly);
     48   }
     49 
     50   void reserveAllocationSpace(uintptr_t CodeSize, uintptr_t DataSizeRO,
     51                               uintptr_t DataSizeRW) override {
     52     return ClientMM->reserveAllocationSpace(CodeSize, DataSizeRO, DataSizeRW);
     53   }
     54 
     55   bool needsToReserveAllocationSpace() override {
     56     return ClientMM->needsToReserveAllocationSpace();
     57   }
     58 
     59   void notifyObjectLoaded(ExecutionEngine *EE,
     60                           const ObjectImage *Obj) override {
     61     ClientMM->notifyObjectLoaded(EE, Obj);
     62   }
     63 
     64   void registerEHFrames(uint8_t *Addr, uint64_t LoadAddr,
     65                         size_t Size) override {
     66     ClientMM->registerEHFrames(Addr, LoadAddr, Size);
     67   }
     68 
     69   void deregisterEHFrames(uint8_t *Addr, uint64_t LoadAddr,
     70                           size_t Size) override {
     71     ClientMM->deregisterEHFrames(Addr, LoadAddr, Size);
     72   }
     73 
     74   bool finalizeMemory(std::string *ErrMsg = nullptr) override {
     75     return ClientMM->finalizeMemory(ErrMsg);
     76   }
     77 
     78 private:
     79   MCJIT *ParentEngine;
     80   std::unique_ptr<RTDyldMemoryManager> ClientMM;
     81 };
     82 
     83 // About Module states: added->loaded->finalized.
     84 //
     85 // The purpose of the "added" state is having modules in standby. (added=known
     86 // but not compiled). The idea is that you can add a module to provide function
     87 // definitions but if nothing in that module is referenced by a module in which
     88 // a function is executed (note the wording here because it's not exactly the
     89 // ideal case) then the module never gets compiled. This is sort of lazy
     90 // compilation.
     91 //
     92 // The purpose of the "loaded" state (loaded=compiled and required sections
     93 // copied into local memory but not yet ready for execution) is to have an
     94 // intermediate state wherein clients can remap the addresses of sections, using
     95 // MCJIT::mapSectionAddress, (in preparation for later copying to a new location
     96 // or an external process) before relocations and page permissions are applied.
     97 //
     98 // It might not be obvious at first glance, but the "remote-mcjit" case in the
     99 // lli tool does this.  In that case, the intermediate action is taken by the
    100 // RemoteMemoryManager in response to the notifyObjectLoaded function being
    101 // called.
    102 
    103 class MCJIT : public ExecutionEngine {
    104   MCJIT(Module *M, TargetMachine *tm, RTDyldMemoryManager *MemMgr,
    105         bool AllocateGVsWithCode);
    106 
    107   typedef llvm::SmallPtrSet<Module *, 4> ModulePtrSet;
    108 
    109   class OwningModuleContainer {
    110   public:
    111     OwningModuleContainer() {
    112     }
    113     ~OwningModuleContainer() {
    114       freeModulePtrSet(AddedModules);
    115       freeModulePtrSet(LoadedModules);
    116       freeModulePtrSet(FinalizedModules);
    117     }
    118 
    119     ModulePtrSet::iterator begin_added() { return AddedModules.begin(); }
    120     ModulePtrSet::iterator end_added() { return AddedModules.end(); }
    121 
    122     ModulePtrSet::iterator begin_loaded() { return LoadedModules.begin(); }
    123     ModulePtrSet::iterator end_loaded() { return LoadedModules.end(); }
    124 
    125     ModulePtrSet::iterator begin_finalized() { return FinalizedModules.begin(); }
    126     ModulePtrSet::iterator end_finalized() { return FinalizedModules.end(); }
    127 
    128     void addModule(Module *M) {
    129       AddedModules.insert(M);
    130     }
    131 
    132     bool removeModule(Module *M) {
    133       return AddedModules.erase(M) || LoadedModules.erase(M) ||
    134              FinalizedModules.erase(M);
    135     }
    136 
    137     bool hasModuleBeenAddedButNotLoaded(Module *M) {
    138       return AddedModules.count(M) != 0;
    139     }
    140 
    141     bool hasModuleBeenLoaded(Module *M) {
    142       // If the module is in either the "loaded" or "finalized" sections it
    143       // has been loaded.
    144       return (LoadedModules.count(M) != 0 ) || (FinalizedModules.count(M) != 0);
    145     }
    146 
    147     bool hasModuleBeenFinalized(Module *M) {
    148       return FinalizedModules.count(M) != 0;
    149     }
    150 
    151     bool ownsModule(Module* M) {
    152       return (AddedModules.count(M) != 0) || (LoadedModules.count(M) != 0) ||
    153              (FinalizedModules.count(M) != 0);
    154     }
    155 
    156     void markModuleAsLoaded(Module *M) {
    157       // This checks against logic errors in the MCJIT implementation.
    158       // This function should never be called with either a Module that MCJIT
    159       // does not own or a Module that has already been loaded and/or finalized.
    160       assert(AddedModules.count(M) &&
    161              "markModuleAsLoaded: Module not found in AddedModules");
    162 
    163       // Remove the module from the "Added" set.
    164       AddedModules.erase(M);
    165 
    166       // Add the Module to the "Loaded" set.
    167       LoadedModules.insert(M);
    168     }
    169 
    170     void markModuleAsFinalized(Module *M) {
    171       // This checks against logic errors in the MCJIT implementation.
    172       // This function should never be called with either a Module that MCJIT
    173       // does not own, a Module that has not been loaded or a Module that has
    174       // already been finalized.
    175       assert(LoadedModules.count(M) &&
    176              "markModuleAsFinalized: Module not found in LoadedModules");
    177 
    178       // Remove the module from the "Loaded" section of the list.
    179       LoadedModules.erase(M);
    180 
    181       // Add the Module to the "Finalized" section of the list by inserting it
    182       // before the 'end' iterator.
    183       FinalizedModules.insert(M);
    184     }
    185 
    186     void markAllLoadedModulesAsFinalized() {
    187       for (ModulePtrSet::iterator I = LoadedModules.begin(),
    188                                   E = LoadedModules.end();
    189            I != E; ++I) {
    190         Module *M = *I;
    191         FinalizedModules.insert(M);
    192       }
    193       LoadedModules.clear();
    194     }
    195 
    196   private:
    197     ModulePtrSet AddedModules;
    198     ModulePtrSet LoadedModules;
    199     ModulePtrSet FinalizedModules;
    200 
    201     void freeModulePtrSet(ModulePtrSet& MPS) {
    202       // Go through the module set and delete everything.
    203       for (ModulePtrSet::iterator I = MPS.begin(), E = MPS.end(); I != E; ++I) {
    204         Module *M = *I;
    205         delete M;
    206       }
    207       MPS.clear();
    208     }
    209   };
    210 
    211   TargetMachine *TM;
    212   MCContext *Ctx;
    213   LinkingMemoryManager MemMgr;
    214   RuntimeDyld Dyld;
    215   SmallVector<JITEventListener*, 2> EventListeners;
    216 
    217   OwningModuleContainer OwnedModules;
    218 
    219   SmallVector<object::Archive*, 2> Archives;
    220 
    221   typedef SmallVector<ObjectImage *, 2> LoadedObjectList;
    222   LoadedObjectList  LoadedObjects;
    223 
    224   // An optional ObjectCache to be notified of compiled objects and used to
    225   // perform lookup of pre-compiled code to avoid re-compilation.
    226   ObjectCache *ObjCache;
    227 
    228   Function *FindFunctionNamedInModulePtrSet(const char *FnName,
    229                                             ModulePtrSet::iterator I,
    230                                             ModulePtrSet::iterator E);
    231 
    232   void runStaticConstructorsDestructorsInModulePtrSet(bool isDtors,
    233                                                       ModulePtrSet::iterator I,
    234                                                       ModulePtrSet::iterator E);
    235 
    236 public:
    237   ~MCJIT();
    238 
    239   /// @name ExecutionEngine interface implementation
    240   /// @{
    241   void addModule(Module *M) override;
    242   void addObjectFile(std::unique_ptr<object::ObjectFile> O) override;
    243   void addArchive(object::Archive *O) override;
    244   bool removeModule(Module *M) override;
    245 
    246   /// FindFunctionNamed - Search all of the active modules to find the one that
    247   /// defines FnName.  This is very slow operation and shouldn't be used for
    248   /// general code.
    249   Function *FindFunctionNamed(const char *FnName) override;
    250 
    251   /// Sets the object manager that MCJIT should use to avoid compilation.
    252   void setObjectCache(ObjectCache *manager) override;
    253 
    254   void setProcessAllSections(bool ProcessAllSections) override {
    255     Dyld.setProcessAllSections(ProcessAllSections);
    256   }
    257 
    258   void generateCodeForModule(Module *M) override;
    259 
    260   /// finalizeObject - ensure the module is fully processed and is usable.
    261   ///
    262   /// It is the user-level function for completing the process of making the
    263   /// object usable for execution. It should be called after sections within an
    264   /// object have been relocated using mapSectionAddress.  When this method is
    265   /// called the MCJIT execution engine will reapply relocations for a loaded
    266   /// object.
    267   /// Is it OK to finalize a set of modules, add modules and finalize again.
    268   // FIXME: Do we really need both of these?
    269   void finalizeObject() override;
    270   virtual void finalizeModule(Module *);
    271   void finalizeLoadedModules();
    272 
    273   /// runStaticConstructorsDestructors - This method is used to execute all of
    274   /// the static constructors or destructors for a program.
    275   ///
    276   /// \param isDtors - Run the destructors instead of constructors.
    277   void runStaticConstructorsDestructors(bool isDtors) override;
    278 
    279   void *getPointerToBasicBlock(BasicBlock *BB) override;
    280 
    281   void *getPointerToFunction(Function *F) override;
    282 
    283   void *recompileAndRelinkFunction(Function *F) override;
    284 
    285   void freeMachineCodeForFunction(Function *F) override;
    286 
    287   GenericValue runFunction(Function *F,
    288                            const std::vector<GenericValue> &ArgValues) override;
    289 
    290   /// getPointerToNamedFunction - This method returns the address of the
    291   /// specified function by using the dlsym function call.  As such it is only
    292   /// useful for resolving library symbols, not code generated symbols.
    293   ///
    294   /// If AbortOnFailure is false and no function with the given name is
    295   /// found, this function silently returns a null pointer. Otherwise,
    296   /// it prints a message to stderr and aborts.
    297   ///
    298   void *getPointerToNamedFunction(const std::string &Name,
    299                                   bool AbortOnFailure = true) override;
    300 
    301   /// mapSectionAddress - map a section to its target address space value.
    302   /// Map the address of a JIT section as returned from the memory manager
    303   /// to the address in the target process as the running code will see it.
    304   /// This is the address which will be used for relocation resolution.
    305   void mapSectionAddress(const void *LocalAddress,
    306                          uint64_t TargetAddress) override {
    307     Dyld.mapSectionAddress(LocalAddress, TargetAddress);
    308   }
    309   void RegisterJITEventListener(JITEventListener *L) override;
    310   void UnregisterJITEventListener(JITEventListener *L) override;
    311 
    312   // If successful, these function will implicitly finalize all loaded objects.
    313   // To get a function address within MCJIT without causing a finalize, use
    314   // getSymbolAddress.
    315   uint64_t getGlobalValueAddress(const std::string &Name) override;
    316   uint64_t getFunctionAddress(const std::string &Name) override;
    317 
    318   TargetMachine *getTargetMachine() override { return TM; }
    319 
    320   /// @}
    321   /// @name (Private) Registration Interfaces
    322   /// @{
    323 
    324   static void Register() {
    325     MCJITCtor = createJIT;
    326   }
    327 
    328   static ExecutionEngine *createJIT(Module *M,
    329                                     std::string *ErrorStr,
    330                                     RTDyldMemoryManager *MemMgr,
    331                                     bool GVsWithCode,
    332                                     TargetMachine *TM);
    333 
    334   // @}
    335 
    336   // This is not directly exposed via the ExecutionEngine API, but it is
    337   // used by the LinkingMemoryManager.
    338   uint64_t getSymbolAddress(const std::string &Name,
    339                           bool CheckFunctionsOnly);
    340 
    341 protected:
    342   /// emitObject -- Generate a JITed object in memory from the specified module
    343   /// Currently, MCJIT only supports a single module and the module passed to
    344   /// this function call is expected to be the contained module.  The module
    345   /// is passed as a parameter here to prepare for multiple module support in
    346   /// the future.
    347   ObjectBufferStream* emitObject(Module *M);
    348 
    349   void NotifyObjectEmitted(const ObjectImage& Obj);
    350   void NotifyFreeingObject(const ObjectImage& Obj);
    351 
    352   uint64_t getExistingSymbolAddress(const std::string &Name);
    353   Module *findModuleForSymbol(const std::string &Name,
    354                               bool CheckFunctionsOnly);
    355 };
    356 
    357 } // End llvm namespace
    358 
    359 #endif
    360