JIT.cpp revision 4fe16d607d11e29d742208894909733f5ad01f8f 
1//===-- JIT.cpp - LLVM Just in Time Compiler ------------------------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file was developed by the LLVM research group and is distributed under
6// the University of Illinois Open Source License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This tool implements a just-in-time compiler for LLVM, allowing direct
11// execution of LLVM bytecode in an efficient manner.
12//
13//===----------------------------------------------------------------------===//
14
15#include "JIT.h"
16#include "llvm/Constants.h"
17#include "llvm/DerivedTypes.h"
18#include "llvm/Function.h"
19#include "llvm/GlobalVariable.h"
20#include "llvm/Instructions.h"
21#include "llvm/ModuleProvider.h"
22#include "llvm/CodeGen/MachineCodeEmitter.h"
23#include "llvm/CodeGen/MachineFunction.h"
24#include "llvm/ExecutionEngine/GenericValue.h"
25#include "llvm/Support/MutexGuard.h"
26#include "llvm/System/DynamicLibrary.h"
27#include "llvm/Target/TargetData.h"
28#include "llvm/Target/TargetMachine.h"
29#include "llvm/Target/TargetJITInfo.h"
30using namespace llvm;
31
32#ifdef __APPLE__
33#include <AvailabilityMacros.h>
34#if defined(MAC_OS_X_VERSION_10_4) && \
35    ((MAC_OS_X_VERSION_MIN_REQUIRED > MAC_OS_X_VERSION_10_4) || \
36     (MAC_OS_X_VERSION_MIN_REQUIRED == MAC_OS_X_VERSION_10_4 && \
37      __APPLE_CC__ >= 5330))
38// __dso_handle is resolved by Mac OS X dynamic linker.
39extern void *__dso_handle __attribute__ ((__visibility__ ("hidden")));
40#endif
41#endif
42
43static struct RegisterJIT {
44  RegisterJIT() { JIT::Register(); }
45} JITRegistrator;
46
47namespace llvm {
48  void LinkInJIT() {
49  }
50}
51
52JIT::JIT(ModuleProvider *MP, TargetMachine &tm, TargetJITInfo &tji)
53  : ExecutionEngine(MP), TM(tm), TJI(tji), state(MP) {
54  setTargetData(TM.getTargetData());
55
56  // Initialize MCE
57  MCE = createEmitter(*this);
58
59  // Add target data
60  MutexGuard locked(lock);
61  FunctionPassManager &PM = state.getPM(locked);
62  PM.add(new TargetData(*TM.getTargetData()));
63
64  // Turn the machine code intermediate representation into bytes in memory that
65  // may be executed.
66  if (TM.addPassesToEmitMachineCode(PM, *MCE, false /*fast*/)) {
67    cerr << "Target does not support machine code emission!\n";
68    abort();
69  }
70
71  // Initialize passes.
72  PM.doInitialization();
73}
74
75JIT::~JIT() {
76  delete MCE;
77  delete &TM;
78}
79
80/// run - Start execution with the specified function and arguments.
81///
82GenericValue JIT::runFunction(Function *F,
83                              const std::vector<GenericValue> &ArgValues) {
84  assert(F && "Function *F was null at entry to run()");
85
86  void *FPtr = getPointerToFunction(F);
87  assert(FPtr && "Pointer to fn's code was null after getPointerToFunction");
88  const FunctionType *FTy = F->getFunctionType();
89  const Type *RetTy = FTy->getReturnType();
90
91  assert((FTy->getNumParams() <= ArgValues.size() || FTy->isVarArg()) &&
92         "Too many arguments passed into function!");
93  assert(FTy->getNumParams() == ArgValues.size() &&
94         "This doesn't support passing arguments through varargs (yet)!");
95
96  // Handle some common cases first.  These cases correspond to common `main'
97  // prototypes.
98  if (RetTy == Type::Int32Ty || RetTy == Type::Int32Ty || RetTy == Type::VoidTy) {
99    switch (ArgValues.size()) {
100    case 3:
101      if ((FTy->getParamType(0) == Type::Int32Ty ||
102           FTy->getParamType(0) == Type::Int32Ty) &&
103          isa<PointerType>(FTy->getParamType(1)) &&
104          isa<PointerType>(FTy->getParamType(2))) {
105        int (*PF)(int, char **, const char **) =
106          (int(*)(int, char **, const char **))(intptr_t)FPtr;
107
108        // Call the function.
109        GenericValue rv;
110        rv.Int32Val = PF(ArgValues[0].Int32Val, (char **)GVTOP(ArgValues[1]),
111                       (const char **)GVTOP(ArgValues[2]));
112        return rv;
113      }
114      break;
115    case 2:
116      if ((FTy->getParamType(0) == Type::Int32Ty ||
117           FTy->getParamType(0) == Type::Int32Ty) &&
118          isa<PointerType>(FTy->getParamType(1))) {
119        int (*PF)(int, char **) = (int(*)(int, char **))(intptr_t)FPtr;
120
121        // Call the function.
122        GenericValue rv;
123        rv.Int32Val = PF(ArgValues[0].Int32Val, (char **)GVTOP(ArgValues[1]));
124        return rv;
125      }
126      break;
127    case 1:
128      if (FTy->getNumParams() == 1 &&
129          (FTy->getParamType(0) == Type::Int32Ty ||
130           FTy->getParamType(0) == Type::Int32Ty)) {
131        GenericValue rv;
132        int (*PF)(int) = (int(*)(int))(intptr_t)FPtr;
133        rv.Int32Val = PF(ArgValues[0].Int32Val);
134        return rv;
135      }
136      break;
137    }
138  }
139
140  // Handle cases where no arguments are passed first.
141  if (ArgValues.empty()) {
142    GenericValue rv;
143    switch (RetTy->getTypeID()) {
144    default: assert(0 && "Unknown return type for function call!");
145    case Type::Int1TyID:
146      rv.Int1Val = ((bool(*)())(intptr_t)FPtr)();
147      return rv;
148    case Type::Int8TyID:
149      rv.Int8Val = ((char(*)())(intptr_t)FPtr)();
150      return rv;
151    case Type::Int16TyID:
152      rv.Int16Val = ((short(*)())(intptr_t)FPtr)();
153      return rv;
154    case Type::VoidTyID:
155    case Type::Int32TyID:
156      rv.Int32Val = ((int(*)())(intptr_t)FPtr)();
157      return rv;
158    case Type::Int64TyID:
159      rv.Int64Val = ((int64_t(*)())(intptr_t)FPtr)();
160      return rv;
161    case Type::FloatTyID:
162      rv.FloatVal = ((float(*)())(intptr_t)FPtr)();
163      return rv;
164    case Type::DoubleTyID:
165      rv.DoubleVal = ((double(*)())(intptr_t)FPtr)();
166      return rv;
167    case Type::PointerTyID:
168      return PTOGV(((void*(*)())(intptr_t)FPtr)());
169    }
170  }
171
172  // Okay, this is not one of our quick and easy cases.  Because we don't have a
173  // full FFI, we have to codegen a nullary stub function that just calls the
174  // function we are interested in, passing in constants for all of the
175  // arguments.  Make this function and return.
176
177  // First, create the function.
178  FunctionType *STy=FunctionType::get(RetTy, std::vector<const Type*>(), false);
179  Function *Stub = new Function(STy, Function::InternalLinkage, "",
180                                F->getParent());
181
182  // Insert a basic block.
183  BasicBlock *StubBB = new BasicBlock("", Stub);
184
185  // Convert all of the GenericValue arguments over to constants.  Note that we
186  // currently don't support varargs.
187  std::vector<Value*> Args;
188  for (unsigned i = 0, e = ArgValues.size(); i != e; ++i) {
189    Constant *C = 0;
190    const Type *ArgTy = FTy->getParamType(i);
191    const GenericValue &AV = ArgValues[i];
192    switch (ArgTy->getTypeID()) {
193    default: assert(0 && "Unknown argument type for function call!");
194    case Type::Int1TyID:   C = ConstantInt::get(AV.Int1Val); break;
195    case Type::Int8TyID:   C = ConstantInt::get(ArgTy, AV.Int8Val);  break;
196    case Type::Int16TyID:  C = ConstantInt::get(ArgTy, AV.Int16Val);  break;
197    case Type::Int32TyID:  C = ConstantInt::get(ArgTy, AV.Int32Val);    break;
198    case Type::Int64TyID:  C = ConstantInt::get(ArgTy, AV.Int64Val);   break;
199    case Type::FloatTyID:  C = ConstantFP ::get(ArgTy, AV.FloatVal);  break;
200    case Type::DoubleTyID: C = ConstantFP ::get(ArgTy, AV.DoubleVal); break;
201    case Type::PointerTyID:
202      void *ArgPtr = GVTOP(AV);
203      if (sizeof(void*) == 4) {
204        C = ConstantInt::get(Type::Int32Ty, (int)(intptr_t)ArgPtr);
205      } else {
206        C = ConstantInt::get(Type::Int64Ty, (intptr_t)ArgPtr);
207      }
208      C = ConstantExpr::getIntToPtr(C, ArgTy);  // Cast the integer to pointer
209      break;
210    }
211    Args.push_back(C);
212  }
213
214  CallInst *TheCall = new CallInst(F, Args, "", StubBB);
215  TheCall->setTailCall();
216  if (TheCall->getType() != Type::VoidTy)
217    new ReturnInst(TheCall, StubBB);             // Return result of the call.
218  else
219    new ReturnInst(StubBB);                      // Just return void.
220
221  // Finally, return the value returned by our nullary stub function.
222  return runFunction(Stub, std::vector<GenericValue>());
223}
224
225/// runJITOnFunction - Run the FunctionPassManager full of
226/// just-in-time compilation passes on F, hopefully filling in
227/// GlobalAddress[F] with the address of F's machine code.
228///
229void JIT::runJITOnFunction(Function *F) {
230  static bool isAlreadyCodeGenerating = false;
231  assert(!isAlreadyCodeGenerating && "Error: Recursive compilation detected!");
232
233  MutexGuard locked(lock);
234
235  // JIT the function
236  isAlreadyCodeGenerating = true;
237  state.getPM(locked).run(*F);
238  isAlreadyCodeGenerating = false;
239
240  // If the function referred to a global variable that had not yet been
241  // emitted, it allocates memory for the global, but doesn't emit it yet.  Emit
242  // all of these globals now.
243  while (!state.getPendingGlobals(locked).empty()) {
244    const GlobalVariable *GV = state.getPendingGlobals(locked).back();
245    state.getPendingGlobals(locked).pop_back();
246    EmitGlobalVariable(GV);
247  }
248}
249
250/// getPointerToFunction - This method is used to get the address of the
251/// specified function, compiling it if neccesary.
252///
253void *JIT::getPointerToFunction(Function *F) {
254  MutexGuard locked(lock);
255
256  if (void *Addr = getPointerToGlobalIfAvailable(F))
257    return Addr;   // Check if function already code gen'd
258
259  // Make sure we read in the function if it exists in this Module.
260  if (F->hasNotBeenReadFromBytecode()) {
261    // Determine the module provider this function is provided by.
262    Module *M = F->getParent();
263    ModuleProvider *MP = 0;
264    for (unsigned i = 0, e = Modules.size(); i != e; ++i) {
265      if (Modules[i]->getModule() == M) {
266        MP = Modules[i];
267        break;
268      }
269    }
270    assert(MP && "Function isn't in a module we know about!");
271
272    std::string ErrorMsg;
273    if (MP->materializeFunction(F, &ErrorMsg)) {
274      cerr << "Error reading function '" << F->getName()
275           << "' from bytecode file: " << ErrorMsg << "\n";
276      abort();
277    }
278  }
279
280  if (F->isExternal()) {
281    void *Addr = getPointerToNamedFunction(F->getName());
282    addGlobalMapping(F, Addr);
283    return Addr;
284  }
285
286  runJITOnFunction(F);
287
288  void *Addr = getPointerToGlobalIfAvailable(F);
289  assert(Addr && "Code generation didn't add function to GlobalAddress table!");
290  return Addr;
291}
292
293/// getOrEmitGlobalVariable - Return the address of the specified global
294/// variable, possibly emitting it to memory if needed.  This is used by the
295/// Emitter.
296void *JIT::getOrEmitGlobalVariable(const GlobalVariable *GV) {
297  MutexGuard locked(lock);
298
299  void *Ptr = getPointerToGlobalIfAvailable(GV);
300  if (Ptr) return Ptr;
301
302  // If the global is external, just remember the address.
303  if (GV->isExternal()) {
304#if defined(__APPLE__) && defined(MAC_OS_X_VERSION_10_4) && \
305    ((MAC_OS_X_VERSION_MIN_REQUIRED > MAC_OS_X_VERSION_10_4) || \
306     (MAC_OS_X_VERSION_MIN_REQUIRED == MAC_OS_X_VERSION_10_4 && \
307      __APPLE_CC__ >= 5330))
308    // Apple gcc defaults to -fuse-cxa-atexit (i.e. calls __cxa_atexit instead
309    // of atexit). It passes the address of linker generated symbol __dso_handle
310    // to the function.
311    // This configuration change happened at version 5330.
312    if (GV->getName() == "__dso_handle")
313      return (void*)&__dso_handle;
314#endif
315    Ptr = sys::DynamicLibrary::SearchForAddressOfSymbol(GV->getName().c_str());
316    if (Ptr == 0) {
317      cerr << "Could not resolve external global address: "
318           << GV->getName() << "\n";
319      abort();
320    }
321  } else {
322    // If the global hasn't been emitted to memory yet, allocate space.  We will
323    // actually initialize the global after current function has finished
324    // compilation.
325    const Type *GlobalType = GV->getType()->getElementType();
326    size_t S = getTargetData()->getTypeSize(GlobalType);
327    size_t A = getTargetData()->getTypeAlignment(GlobalType);
328    if (A <= 8) {
329      Ptr = malloc(S);
330    } else {
331      // Allocate S+A bytes of memory, then use an aligned pointer within that
332      // space.
333      Ptr = malloc(S+A);
334      unsigned MisAligned = ((intptr_t)Ptr & (A-1));
335      Ptr = (char*)Ptr + (MisAligned ? (A-MisAligned) : 0);
336    }
337    state.getPendingGlobals(locked).push_back(GV);
338  }
339  addGlobalMapping(GV, Ptr);
340  return Ptr;
341}
342
343
344/// recompileAndRelinkFunction - This method is used to force a function
345/// which has already been compiled, to be compiled again, possibly
346/// after it has been modified. Then the entry to the old copy is overwritten
347/// with a branch to the new copy. If there was no old copy, this acts
348/// just like JIT::getPointerToFunction().
349///
350void *JIT::recompileAndRelinkFunction(Function *F) {
351  void *OldAddr = getPointerToGlobalIfAvailable(F);
352
353  // If it's not already compiled there is no reason to patch it up.
354  if (OldAddr == 0) { return getPointerToFunction(F); }
355
356  // Delete the old function mapping.
357  addGlobalMapping(F, 0);
358
359  // Recodegen the function
360  runJITOnFunction(F);
361
362  // Update state, forward the old function to the new function.
363  void *Addr = getPointerToGlobalIfAvailable(F);
364  assert(Addr && "Code generation didn't add function to GlobalAddress table!");
365  TJI.replaceMachineCodeForFunction(OldAddr, Addr);
366  return Addr;
367}
368
369