Initial checkin of new LLI with JIT compiler
authorChris Lattner <sabre@nondot.org>
Tue, 24 Dec 2002 00:01:05 +0000 (00:01 +0000)
committerChris Lattner <sabre@nondot.org>
Tue, 24 Dec 2002 00:01:05 +0000 (00:01 +0000)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@5126 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/ExecutionEngine/ExecutionEngine.h [new file with mode: 0644]
include/llvm/ExecutionEngine/GenericValue.h [new file with mode: 0644]
lib/ExecutionEngine/ExecutionEngine.cpp [new file with mode: 0644]
lib/ExecutionEngine/Interpreter/Interpreter.cpp [new file with mode: 0644]
lib/ExecutionEngine/Interpreter/Makefile [new file with mode: 0644]
lib/ExecutionEngine/JIT/Callback.cpp [new file with mode: 0644]
lib/ExecutionEngine/JIT/GlobalVars.cpp [new file with mode: 0644]
lib/ExecutionEngine/JIT/JIT.cpp [new file with mode: 0644]
lib/ExecutionEngine/JIT/JITEmitter.cpp [new file with mode: 0644]
lib/ExecutionEngine/JIT/Makefile [new file with mode: 0644]
lib/ExecutionEngine/JIT/VM.cpp [new file with mode: 0644]

diff --git a/include/llvm/ExecutionEngine/ExecutionEngine.h b/include/llvm/ExecutionEngine/ExecutionEngine.h
new file mode 100644 (file)
index 0000000..57eee95
--- /dev/null
@@ -0,0 +1,86 @@
+//===- ExecutionEngine.h - Abstract Execution Engine Interface --*- C++ -*-===//
+//
+// This file defines the abstract interface that implements execution support
+// for LLVM.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef EXECUTION_ENGINE_H
+#define EXECUTION_ENGINE_H
+
+#include <vector>
+#include <string>
+#include <map>
+class Constant;
+class Type;
+class GlobalValue;
+class Function;
+class Module;
+class TargetData;
+union GenericValue;
+
+class ExecutionEngine {
+  Module &CurMod;
+  const TargetData *TD;
+
+protected:
+  // GlobalAddress - A mapping between LLVM global values and their actualized
+  // version...
+  std::map<const GlobalValue*, void *> GlobalAddress;
+
+  void setTargetData(const TargetData &td) {
+    TD = &td;
+    emitGlobals();
+  }
+public:
+  ExecutionEngine(Module *M) : CurMod(*M) {
+    assert(M && "Module is null?");
+  }
+  virtual ~ExecutionEngine();
+  
+  Module &getModule() const { return CurMod; }
+  const TargetData &getTargetData() const { return *TD; }
+
+  /// run - Start execution with the specified function and arguments.
+  ///
+  virtual int run(const std::string &FnName,
+                 const std::vector<std::string> &Args) = 0;
+
+  /// createJIT - Create an return a new JIT compiler if there is one available
+  /// for the current target.  Otherwise it returns null.
+  ///
+  static ExecutionEngine *createJIT(Module *M, unsigned Config);
+
+  /// createInterpreter - Create a new interpreter object.  This can never fail.
+  ///
+  static ExecutionEngine *createInterpreter(Module *M, unsigned Config,
+                                           bool DebugMode, bool TraceMode);
+
+  void addGlobalMapping(const Function *F, void *Addr) {
+    void *&CurVal = GlobalAddress[(const GlobalValue*)F];
+    assert(CurVal == 0 && "GlobalMapping already established!");
+    CurVal = Addr;
+  }
+
+  // getPointerToGlobal - This returns the address of the specified global
+  // value.  This may involve code generation if it's a function.
+  //
+  void *getPointerToGlobal(const GlobalValue *GV);
+
+  // getPointerToFunction - The different EE's represent function bodies in
+  // different ways.  They should each implement this to say what a function
+  // pointer should look like.
+  //
+  virtual void *getPointerToFunction(const Function *F) = 0;
+
+private:
+  void emitGlobals();
+
+public:   // FIXME: protected:   // API shared among subclasses
+  GenericValue getConstantValue(const Constant *C);
+  void StoreValueToMemory(GenericValue Val, GenericValue *Ptr, const Type *Ty);
+  void *CreateArgv(const std::vector<std::string> &InputArgv);
+  void InitializeMemory(const Constant *Init, void *Addr);
+};
+
+#endif
diff --git a/include/llvm/ExecutionEngine/GenericValue.h b/include/llvm/ExecutionEngine/GenericValue.h
new file mode 100644 (file)
index 0000000..f5e70e5
--- /dev/null
@@ -0,0 +1,40 @@
+//===-- GenericValue.h - Represent any type of LLVM value -------*- C++ -*-===//
+// 
+// The GenericValue class is used to represent an LLVM value of arbitrary type.
+//
+//===----------------------------------------------------------------------===//
+
+
+#ifndef GENERIC_VALUE_H
+#define GENERIC_VALUE_H
+
+#include "Support/DataTypes.h"
+
+typedef uint64_t PointerTy;
+
+union GenericValue {
+  bool            BoolVal;
+  unsigned char   UByteVal;
+  signed   char   SByteVal;
+  unsigned short  UShortVal;
+  signed   short  ShortVal;
+  unsigned int    UIntVal;
+  signed   int    IntVal;
+  uint64_t        ULongVal;
+  int64_t         LongVal;
+  double          DoubleVal;
+  float           FloatVal;
+  PointerTy       PointerVal;
+  unsigned char   Untyped[8];
+
+  GenericValue() {}
+  GenericValue(void *V) {
+    PointerVal = (PointerTy)(intptr_t)V;
+  }
+};
+
+inline GenericValue PTOGV(void *P) { return GenericValue(P); }
+inline void* GVTOP(const GenericValue &GV) {
+  return (void*)(intptr_t)GV.PointerVal;
+}
+#endif
diff --git a/lib/ExecutionEngine/ExecutionEngine.cpp b/lib/ExecutionEngine/ExecutionEngine.cpp
new file mode 100644 (file)
index 0000000..996afe4
--- /dev/null
@@ -0,0 +1,226 @@
+//===-- ExecutionEngine.cpp - Common Implementation shared by EE's --------===//
+// 
+// This file defines the common interface used by the various execution engine
+// subclasses.
+//
+//===----------------------------------------------------------------------===//
+
+#include "ExecutionEngine.h"
+#include "GenericValue.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Constants.h"
+#include "llvm/Module.h"
+#include "llvm/Target/TargetData.h"
+#include "Support/Statistic.h"
+
+Statistic<> NumInitBytes("lli", "Number of bytes of global vars initialized");
+
+// getPointerToGlobal - This returns the address of the specified global
+// value.  This may involve code generation if it's a function.
+//
+void *ExecutionEngine::getPointerToGlobal(const GlobalValue *GV) {
+  if (const Function *F = dyn_cast<Function>(GV))
+    return getPointerToFunction(F);
+
+  assert(GlobalAddress[GV] && "Global hasn't had an address allocated yet?");
+  return GlobalAddress[GV];
+}
+
+
+GenericValue ExecutionEngine::getConstantValue(const Constant *C) {
+  GenericValue Result;
+#define GET_CONST_VAL(TY, CLASS) \
+  case Type::TY##TyID: Result.TY##Val = cast<CLASS>(C)->getValue(); break
+
+  switch (C->getType()->getPrimitiveID()) {
+    GET_CONST_VAL(Bool   , ConstantBool);
+    GET_CONST_VAL(UByte  , ConstantUInt);
+    GET_CONST_VAL(SByte  , ConstantSInt);
+    GET_CONST_VAL(UShort , ConstantUInt);
+    GET_CONST_VAL(Short  , ConstantSInt);
+    GET_CONST_VAL(UInt   , ConstantUInt);
+    GET_CONST_VAL(Int    , ConstantSInt);
+    GET_CONST_VAL(ULong  , ConstantUInt);
+    GET_CONST_VAL(Long   , ConstantSInt);
+    GET_CONST_VAL(Float  , ConstantFP);
+    GET_CONST_VAL(Double , ConstantFP);
+#undef GET_CONST_VAL
+  case Type::PointerTyID:
+    if (isa<ConstantPointerNull>(C)) {
+      Result.PointerVal = 0;
+    } else if (const ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(C)){
+      Result = PTOGV(getPointerToGlobal(CPR->getValue()));
+
+    } else {
+      assert(0 && "Unknown constant pointer type!");
+    }
+    break;
+  default:
+    cout << "ERROR: Constant unimp for type: " << C->getType() << "\n";
+  }
+  return Result;
+}
+
+void ExecutionEngine::StoreValueToMemory(GenericValue Val, GenericValue *Ptr,
+                                    const Type *Ty) {
+  if (getTargetData().isLittleEndian()) {
+    switch (Ty->getPrimitiveID()) {
+    case Type::BoolTyID:
+    case Type::UByteTyID:
+    case Type::SByteTyID:   Ptr->Untyped[0] = Val.UByteVal; break;
+    case Type::UShortTyID:
+    case Type::ShortTyID:   Ptr->Untyped[0] = Val.UShortVal & 255;
+                            Ptr->Untyped[1] = (Val.UShortVal >> 8) & 255;
+                            break;
+    case Type::FloatTyID:
+    case Type::UIntTyID:
+    case Type::IntTyID:     Ptr->Untyped[0] =  Val.UIntVal        & 255;
+                            Ptr->Untyped[1] = (Val.UIntVal >>  8) & 255;
+                            Ptr->Untyped[2] = (Val.UIntVal >> 16) & 255;
+                            Ptr->Untyped[3] = (Val.UIntVal >> 24) & 255;
+                            break;
+    case Type::DoubleTyID:
+    case Type::ULongTyID:
+    case Type::LongTyID:    
+    case Type::PointerTyID: Ptr->Untyped[0] =  Val.ULongVal        & 255;
+                            Ptr->Untyped[1] = (Val.ULongVal >>  8) & 255;
+                            Ptr->Untyped[2] = (Val.ULongVal >> 16) & 255;
+                            Ptr->Untyped[3] = (Val.ULongVal >> 24) & 255;
+                            Ptr->Untyped[4] = (Val.ULongVal >> 32) & 255;
+                            Ptr->Untyped[5] = (Val.ULongVal >> 40) & 255;
+                            Ptr->Untyped[6] = (Val.ULongVal >> 48) & 255;
+                            Ptr->Untyped[7] = (Val.ULongVal >> 56) & 255;
+                            break;
+    default:
+      cout << "Cannot store value of type " << Ty << "!\n";
+    }
+  } else {
+    switch (Ty->getPrimitiveID()) {
+    case Type::BoolTyID:
+    case Type::UByteTyID:
+    case Type::SByteTyID:   Ptr->Untyped[0] = Val.UByteVal; break;
+    case Type::UShortTyID:
+    case Type::ShortTyID:   Ptr->Untyped[1] = Val.UShortVal & 255;
+                            Ptr->Untyped[0] = (Val.UShortVal >> 8) & 255;
+                            break;
+    case Type::FloatTyID:
+    case Type::UIntTyID:
+    case Type::IntTyID:     Ptr->Untyped[3] =  Val.UIntVal        & 255;
+                            Ptr->Untyped[2] = (Val.UIntVal >>  8) & 255;
+                            Ptr->Untyped[1] = (Val.UIntVal >> 16) & 255;
+                            Ptr->Untyped[0] = (Val.UIntVal >> 24) & 255;
+                            break;
+    case Type::DoubleTyID:
+    case Type::ULongTyID:
+    case Type::LongTyID:    
+    case Type::PointerTyID: Ptr->Untyped[7] =  Val.ULongVal        & 255;
+                            Ptr->Untyped[6] = (Val.ULongVal >>  8) & 255;
+                            Ptr->Untyped[5] = (Val.ULongVal >> 16) & 255;
+                            Ptr->Untyped[4] = (Val.ULongVal >> 24) & 255;
+                            Ptr->Untyped[3] = (Val.ULongVal >> 32) & 255;
+                            Ptr->Untyped[2] = (Val.ULongVal >> 40) & 255;
+                            Ptr->Untyped[1] = (Val.ULongVal >> 48) & 255;
+                            Ptr->Untyped[0] = (Val.ULongVal >> 56) & 255;
+                            break;
+    default:
+      cout << "Cannot store value of type " << Ty << "!\n";
+    }
+  }
+}
+
+// InitializeMemory - Recursive function to apply a Constant value into the
+// specified memory location...
+//
+void ExecutionEngine::InitializeMemory(const Constant *Init, void *Addr) {
+  if (Init->getType()->isFirstClassType()) {
+    GenericValue Val = getConstantValue(Init);
+    StoreValueToMemory(Val, (GenericValue*)Addr, Init->getType());
+    return;
+  }
+
+  switch (Init->getType()->getPrimitiveID()) {
+  case Type::ArrayTyID: {
+    const ConstantArray *CPA = cast<ConstantArray>(Init);
+    const vector<Use> &Val = CPA->getValues();
+    unsigned ElementSize = 
+      getTargetData().getTypeSize(cast<ArrayType>(CPA->getType())->getElementType());
+    for (unsigned i = 0; i < Val.size(); ++i)
+      InitializeMemory(cast<Constant>(Val[i].get()), (char*)Addr+i*ElementSize);
+    return;
+  }
+
+  case Type::StructTyID: {
+    const ConstantStruct *CPS = cast<ConstantStruct>(Init);
+    const StructLayout *SL =
+      getTargetData().getStructLayout(cast<StructType>(CPS->getType()));
+    const vector<Use> &Val = CPS->getValues();
+    for (unsigned i = 0; i < Val.size(); ++i)
+      InitializeMemory(cast<Constant>(Val[i].get()),
+                       (char*)Addr+SL->MemberOffsets[i]);
+    return;
+  }
+
+  default:
+    std::cerr << "Bad Type: " << Init->getType() << "\n";
+    assert(0 && "Unknown constant type to initialize memory with!");
+  }
+}
+
+
+
+void *ExecutionEngine::CreateArgv(const std::vector<std::string> &InputArgv) {
+  // Pointers are 64 bits...
+  PointerTy *Result = new PointerTy[InputArgv.size()+1];  // 64 bit assumption
+  DEBUG(std::cerr << "ARGV = " << (void*)Result << "\n");
+
+  for (unsigned i = 0; i < InputArgv.size(); ++i) {
+    unsigned Size = InputArgv[i].size()+1;
+    char *Dest = new char[Size];
+    DEBUG(std::cerr << "ARGV[" << i << "] = " << (void*)Dest << "\n");
+
+    copy(InputArgv[i].begin(), InputArgv[i].end(), Dest);
+    Dest[Size-1] = 0;
+
+    // Endian safe: Result[i] = (PointerTy)Dest;
+    StoreValueToMemory(PTOGV(Dest), (GenericValue*)(Result+i),
+                       Type::LongTy);  // 64 bit assumption
+  }
+
+  Result[InputArgv.size()] = 0;
+  return Result;
+}
+
+/// EmitGlobals - Emit all of the global variables to memory, storing their
+/// addresses into GlobalAddress.  This must make sure to copy the contents of
+/// their initializers into the memory.
+///
+void ExecutionEngine::emitGlobals() {
+  const TargetData &TD = getTargetData();
+  
+  // Loop over all of the global variables in the program, allocating the memory
+  // to hold them.
+  for (Module::giterator I = getModule().gbegin(), E = getModule().gend();
+       I != E; ++I)
+    if (!I->isExternal()) {
+      // Get the type of the global...
+      const Type *Ty = I->getType()->getElementType();
+      
+      // Allocate some memory for it!
+      unsigned Size = TD.getTypeSize(Ty);
+      GlobalAddress[I] = new char[Size];
+      NumInitBytes += Size;
+
+      DEBUG(std::cerr << "Global '" << I->getName() << "' -> "
+                     << (void*)GlobalAddress[I] << "\n");
+    } else {
+      assert(0 && "References to external globals not handled yet!");
+    }
+  
+  // Now that all of the globals are set up in memory, loop through them all and
+  // initialize their contents.
+  for (Module::giterator I = getModule().gbegin(), E = getModule().gend();
+       I != E; ++I)
+    if (!I->isExternal())
+      InitializeMemory(I->getInitializer(), GlobalAddress[I]);
+}
+
diff --git a/lib/ExecutionEngine/Interpreter/Interpreter.cpp b/lib/ExecutionEngine/Interpreter/Interpreter.cpp
new file mode 100644 (file)
index 0000000..678c77a
--- /dev/null
@@ -0,0 +1,56 @@
+//===- Interpreter.cpp - Top-Level LLVM Interpreter Implementation --------===//
+//
+// This file implements the top-level functionality for the LLVM interpreter.
+// This interpreter is designed to be a very simple, portable, inefficient
+// interpreter.
+//
+//===----------------------------------------------------------------------===//
+
+#include "Interpreter.h"
+#include "llvm/Target/TargetMachineImpls.h"
+
+/// createInterpreter - Create a new interpreter object.  This can never fail.
+///
+ExecutionEngine *ExecutionEngine::createInterpreter(Module *M,
+                                                   unsigned Config,
+                                                   bool DebugMode,
+                                                   bool TraceMode) {
+  return new Interpreter(M, Config, DebugMode, TraceMode);
+}
+
+//===----------------------------------------------------------------------===//
+// Interpreter ctor - Initialize stuff
+//
+Interpreter::Interpreter(Module *M, unsigned Config,
+                        bool DebugMode, bool TraceMode)
+  : ExecutionEngine(M), ExitCode(0), Debug(DebugMode), Trace(TraceMode),
+    CurFrame(-1), TD("lli", (Config & TM::EndianMask) == TM::LittleEndian,
+                    1, 4,
+                    (Config & TM::PtrSizeMask) == TM::PtrSize64 ? 8 : 4,
+                    (Config & TM::PtrSizeMask) == TM::PtrSize64 ? 8 : 4) {
+
+  setTargetData(TD);
+  // Initialize the "backend"
+  initializeExecutionEngine();
+  initializeExternalMethods();
+  CW.setModule(M);  // Update Writer
+}
+
+/// run - Start execution with the specified function and arguments.
+///
+int Interpreter::run(const std::string &MainFunction,
+                    const std::vector<std::string> &Args) {
+  // Start interpreter into the main function...
+  //
+  if (!callMainMethod(MainFunction, Args) && !Debug) {
+    // If not in debug mode and if the call succeeded, run the code now...
+    run();
+  }
+
+  // If debug mode, allow the user to interact... also, if the user pressed 
+  // ctrl-c or execution hit an error, enter the event loop...
+  if (Debug || isStopped())
+    handleUserInput();
+  return ExitCode;
+}
+
diff --git a/lib/ExecutionEngine/Interpreter/Makefile b/lib/ExecutionEngine/Interpreter/Makefile
new file mode 100644 (file)
index 0000000..1a7d3bf
--- /dev/null
@@ -0,0 +1,4 @@
+LEVEL = ../../..
+LIBRARYNAME = lli-interpreter
+
+include $(LEVEL)/Makefile.common
diff --git a/lib/ExecutionEngine/JIT/Callback.cpp b/lib/ExecutionEngine/JIT/Callback.cpp
new file mode 100644 (file)
index 0000000..b843e10
--- /dev/null
@@ -0,0 +1,62 @@
+//===-- Callback.cpp - Trap handler for function resolution ---------------===//
+//
+// This file defines the SIGSEGV handler which is invoked when a reference to a
+// non-codegen'd function is found.
+//
+//===----------------------------------------------------------------------===//
+
+#include "VM.h"
+#include "Support/Statistic.h"
+#include <signal.h>
+#include <ucontext.h>
+#include <iostream>
+
+static VM *TheVM = 0;
+
+static void TrapHandler(int TN, siginfo_t *SI, ucontext_t *ucp) {
+  assert(TN == SIGSEGV && "Should be SIGSEGV!");
+
+#ifdef REG_EIP   /* this code does not compile on Sparc! */
+  if (SI->si_code != SEGV_MAPERR || SI->si_addr != 0 ||
+      ucp->uc_mcontext.gregs[REG_EIP] != 0) {
+    std::cerr << "Bad SEGV encountered!\n";
+    abort();
+  }
+
+  // The call instruction should have pushed the return value onto the stack...
+  unsigned RefAddr = *(unsigned*)ucp->uc_mcontext.gregs[REG_ESP];
+  RefAddr -= 4;  // Backtrack to the reference itself...
+
+  DEBUG(std::cerr << "In SEGV handler! Addr=0x" << std::hex << RefAddr
+                  << " ESP=0x" << ucp->uc_mcontext.gregs[REG_ESP] << std::dec
+                  << ": Resolving call to function: "
+                  << TheVM->getFunctionReferencedName((void*)RefAddr) << "\n");
+
+  // Sanity check to make sure this really is a call instruction...
+  assert(((unsigned char*)RefAddr)[-1] == 0xE8 && "Not a call instr!");
+  
+  unsigned NewVal = (unsigned)TheVM->resolveFunctionReference((void*)RefAddr);
+
+  // Rewrite the call target... so that we don't fault every time we execute
+  // the call.
+  *(unsigned*)RefAddr = NewVal-RefAddr-4;    
+
+  // Change the instruction pointer to be the real target of the call...
+  ucp->uc_mcontext.gregs[REG_EIP] = NewVal;
+
+#endif
+}
+
+
+void VM::registerCallback() {
+  TheVM = this;
+
+  // Register the signal handler...
+  struct sigaction SA;
+  SA.sa_sigaction = (void (*)(int, siginfo_t*, void*))TrapHandler;
+  sigfillset(&SA.sa_mask);               // Block all signals while codegen'ing
+  SA.sa_flags = SA_NOCLDSTOP|SA_SIGINFO; // Get siginfo
+  sigaction(SIGSEGV, &SA, 0);            // Install the handler
+}
+
+
diff --git a/lib/ExecutionEngine/JIT/GlobalVars.cpp b/lib/ExecutionEngine/JIT/GlobalVars.cpp
new file mode 100644 (file)
index 0000000..e69de29
diff --git a/lib/ExecutionEngine/JIT/JIT.cpp b/lib/ExecutionEngine/JIT/JIT.cpp
new file mode 100644 (file)
index 0000000..b2b56a6
--- /dev/null
@@ -0,0 +1,53 @@
+//===-- JIT.cpp - LLVM Just in Time Compiler ------------------------------===//
+//
+// This file implements the top-level support for creating a Just-In-Time
+// compiler for the current architecture.
+//
+//===----------------------------------------------------------------------===//
+
+#include "VM.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetMachineImpls.h"
+#include "llvm/Module.h"
+
+
+/// createJIT - Create an return a new JIT compiler if there is one available
+/// for the current target.  Otherwise it returns null.
+///
+ExecutionEngine *ExecutionEngine::createJIT(Module *M, unsigned Config) {
+  // FIXME: This should be controlled by which subdirectory gets linked in!
+#if !defined(i386) && !defined(__i386__) && !defined(__x86__)
+  return 0;
+#endif
+  // Allocate a target... in the future this will be controllable on the
+  // command line.
+  TargetMachine *Target = allocateX86TargetMachine(Config);
+  assert(Target && "Could not allocate X86 target machine!");
+
+  // Create the virtual machine object...
+  return new VM(M, Target);
+}
+
+VM::VM(Module *M, TargetMachine *tm) : ExecutionEngine(M), TM(*tm) {
+  setTargetData(TM.getTargetData());
+  MCE = createEmitter(*this);  // Initialize MCE
+  setupPassManager();
+  registerCallback();
+}
+
+int VM::run(const std::string &FnName, const std::vector<std::string> &Args) {
+  Function *F = getModule().getNamedFunction(FnName);
+  if (F == 0) {
+    std::cerr << "Could not find function '" << FnName <<"' in module!\n";
+    return 1;
+  }
+
+  int(*PF)(int, char**) = (int(*)(int, char**))getPointerToFunction(F);
+  assert(PF != 0 && "Null pointer to function?");
+
+  // Build an argv vector...
+  char **Argv = (char**)CreateArgv(Args);
+
+  // Call the main function...
+  return PF(Args.size(), Argv);
+}
diff --git a/lib/ExecutionEngine/JIT/JITEmitter.cpp b/lib/ExecutionEngine/JIT/JITEmitter.cpp
new file mode 100644 (file)
index 0000000..253a229
--- /dev/null
@@ -0,0 +1,107 @@
+//===-- Emitter.cpp - Write machine code to executable memory -------------===//
+//
+// This file defines a MachineCodeEmitter object that is used by Jello to write
+// machine code to memory and remember where relocatable values lie.
+//
+//===----------------------------------------------------------------------===//
+
+#include "VM.h"
+#include "llvm/CodeGen/MachineCodeEmitter.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/Function.h"
+#include "Support/Statistic.h"
+
+namespace {
+  Statistic<> NumBytes("jello", "Number of bytes of machine code compiled");
+
+  class Emitter : public MachineCodeEmitter {
+    VM &TheVM;
+
+    unsigned char *CurBlock;
+    unsigned char *CurByte;
+    
+    std::vector<std::pair<BasicBlock*, unsigned *> > BBRefs;
+    std::map<BasicBlock*, unsigned> BBLocations;
+  public:
+    Emitter(VM &vm) : TheVM(vm) {}
+
+    virtual void startFunction(MachineFunction &F);
+    virtual void finishFunction(MachineFunction &F);
+    virtual void startBasicBlock(MachineBasicBlock &BB);
+    virtual void emitByte(unsigned char B);
+    virtual void emitPCRelativeDisp(Value *V);
+    virtual void emitGlobalAddress(GlobalValue *V);
+  };
+}
+
+MachineCodeEmitter *VM::createEmitter(VM &V) {
+  return new Emitter(V);
+}
+
+
+#define _POSIX_MAPPED_FILES
+#include <unistd.h>
+#include <sys/mman.h>
+
+static void *getMemory() {
+  return mmap(0, 4096*2, PROT_READ|PROT_WRITE|PROT_EXEC,
+              MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
+}
+
+
+void Emitter::startFunction(MachineFunction &F) {
+  CurBlock = (unsigned char *)getMemory();
+  CurByte = CurBlock;  // Start writing at the beginning of the fn.
+  TheVM.addGlobalMapping(F.getFunction(), CurBlock);
+}
+
+void Emitter::finishFunction(MachineFunction &F) {
+  for (unsigned i = 0, e = BBRefs.size(); i != e; ++i) {
+    unsigned Location = BBLocations[BBRefs[i].first];
+    unsigned *Ref = BBRefs[i].second;
+    *Ref = Location-(unsigned)Ref-4;
+  }
+  BBRefs.clear();
+  BBLocations.clear();
+
+  NumBytes += CurByte-CurBlock;
+
+  DEBUG(std::cerr << "Finished CodeGen of [" << std::hex << (unsigned)CurBlock
+                  << std::dec << "] Function: " << F.getFunction()->getName()
+                  << ": " << CurByte-CurBlock << " bytes of text\n");
+}
+
+void Emitter::startBasicBlock(MachineBasicBlock &BB) {
+  BBLocations[BB.getBasicBlock()] = (unsigned)CurByte;
+}
+
+
+void Emitter::emitByte(unsigned char B) {
+  *CurByte++ = B;   // Write the byte to memory
+}
+
+
+// emitPCRelativeDisp - For functions, just output a displacement that will
+// cause a reference to the zero page, which will cause a seg-fault, causing
+// things to get resolved on demand.  Keep track of these markers.
+//
+// For basic block references, keep track of where the references are so they
+// may be patched up when the basic block is defined.
+//
+void Emitter::emitPCRelativeDisp(Value *V) {
+  if (Function *F = dyn_cast<Function>(V)) {
+    TheVM.addFunctionRef(CurByte, F);
+    unsigned ZeroAddr = -(unsigned)CurByte-4; // Calculate displacement to null
+    *(unsigned*)CurByte = ZeroAddr;           // 4 byte offset
+    CurByte += 4;
+  } else {
+    BasicBlock *BB = cast<BasicBlock>(V);     // Keep track of reference...
+    BBRefs.push_back(std::make_pair(BB, (unsigned*)CurByte));
+    CurByte += 4;
+  }
+}
+
+void Emitter::emitGlobalAddress(GlobalValue *V) {
+  *(void**)CurByte = TheVM.getPointerToGlobal(V);
+  CurByte += 4;
+}
diff --git a/lib/ExecutionEngine/JIT/Makefile b/lib/ExecutionEngine/JIT/Makefile
new file mode 100644 (file)
index 0000000..bc3f709
--- /dev/null
@@ -0,0 +1,4 @@
+LEVEL = ../../..
+LIBRARYNAME = lli-jit
+
+include $(LEVEL)/Makefile.common
diff --git a/lib/ExecutionEngine/JIT/VM.cpp b/lib/ExecutionEngine/JIT/VM.cpp
new file mode 100644 (file)
index 0000000..f66d4d7
--- /dev/null
@@ -0,0 +1,84 @@
+//===-- jello.cpp - LLVM Just in Time Compiler ----------------------------===//
+//
+// This tool implements a just-in-time compiler for LLVM, allowing direct
+// execution of LLVM bytecode in an efficient manner.
+//
+//===----------------------------------------------------------------------===//
+
+#include "VM.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/CodeGen/MachineCodeEmitter.h"
+#include "llvm/Function.h"
+#include <dlfcn.h>    // dlsym access
+
+
+VM::~VM() {
+  delete MCE;
+  delete &TM;
+}
+
+/// setupPassManager - Initialize the VM PassManager object with all of the
+/// passes needed for the target to generate code.
+///
+void VM::setupPassManager() {
+  // Compile LLVM Code down to machine code in the intermediate representation
+  if (TM.addPassesToJITCompile(PM)) {
+    std::cerr << "lli: target '" << TM.getName()
+              << "' doesn't support JIT compilation!\n";
+    abort();
+  }
+
+  // Turn the machine code intermediate representation into bytes in memory that
+  // may be executed.
+  //
+  if (TM.addPassesToEmitMachineCode(PM, *MCE)) {
+    std::cerr << "lli: target '" << TM.getName()
+              << "' doesn't support machine code emission!\n";
+    abort();
+  }
+}
+
+void *VM::resolveFunctionReference(void *RefAddr) {
+  Function *F = FunctionRefs[RefAddr];
+  assert(F && "Reference address not known!");
+
+  void *Addr = getPointerToFunction(F);
+  assert(Addr && "Pointer to function unknown!");
+
+  FunctionRefs.erase(RefAddr);
+  return Addr;
+}
+
+const std::string &VM::getFunctionReferencedName(void *RefAddr) {
+  return FunctionRefs[RefAddr]->getName();
+}
+
+static void NoopFn() {}
+
+/// getPointerToFunction - This method is used to get the address of the
+/// specified function, compiling it if neccesary.
+///
+void *VM::getPointerToFunction(const Function *F) {
+  void *&Addr = GlobalAddress[F];   // Function already code gen'd
+  if (Addr) return Addr;
+
+  if (F->isExternal()) {
+    // If it's an external function, look it up in the process image...
+    void *Ptr = dlsym(0, F->getName().c_str());
+    if (Ptr == 0) {
+      std::cerr << "WARNING: Cannot resolve fn '" << F->getName()
+                << "' using a dummy noop function instead!\n";
+      Ptr = (void*)NoopFn;
+    }
+
+    return Addr = Ptr;
+  }
+
+  // JIT all of the functions in the module.  Eventually this will JIT functions
+  // on demand.  This has the effect of populating all of the non-external
+  // functions into the GlobalAddress table.
+  PM.run(getModule());
+
+  assert(Addr && "Code generation didn't add function to GlobalAddress table!");
+  return Addr;
+}