JIT.cpp revision 9f2f142d255bc96f109dd5c6524a485937b1f3a1
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 bitcode 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" 30 31#include "llvm/Config/config.h" 32 33using namespace llvm; 34 35#ifdef __APPLE__ 36// Apple gcc defaults to -fuse-cxa-atexit (i.e. calls __cxa_atexit instead 37// of atexit). It passes the address of linker generated symbol __dso_handle 38// to the function. 39// This configuration change happened at version 5330. 40# include <AvailabilityMacros.h> 41# if defined(MAC_OS_X_VERSION_10_4) && \ 42 ((MAC_OS_X_VERSION_MIN_REQUIRED > MAC_OS_X_VERSION_10_4) || \ 43 (MAC_OS_X_VERSION_MIN_REQUIRED == MAC_OS_X_VERSION_10_4 && \ 44 __APPLE_CC__ >= 5330)) 45# ifndef HAVE___DSO_HANDLE 46# define HAVE___DSO_HANDLE 1 47# endif 48# endif 49#endif 50 51#if HAVE___DSO_HANDLE 52extern void *__dso_handle __attribute__ ((__visibility__ ("hidden"))); 53#endif 54 55static struct RegisterJIT { 56 RegisterJIT() { JIT::Register(); } 57} JITRegistrator; 58 59namespace llvm { 60 void LinkInJIT() { 61 } 62} 63 64JIT::JIT(ModuleProvider *MP, TargetMachine &tm, TargetJITInfo &tji) 65 : ExecutionEngine(MP), TM(tm), TJI(tji), jitstate(MP) { 66 setTargetData(TM.getTargetData()); 67 68 // Initialize MCE 69 MCE = createEmitter(*this, 0); 70 71 // Add target data 72 MutexGuard locked(lock); 73 FunctionPassManager &PM = jitstate.getPM(locked); 74 PM.add(new TargetData(*TM.getTargetData())); 75 76 // Turn the machine code intermediate representation into bytes in memory that 77 // may be executed. 78 if (TM.addPassesToEmitMachineCode(PM, *MCE, false /*fast*/)) { 79 cerr << "Target does not support machine code emission!\n"; 80 abort(); 81 } 82 83 // Initialize passes. 84 PM.doInitialization(); 85} 86 87JIT::~JIT() { 88 delete MCE; 89 delete &TM; 90} 91 92/// run - Start execution with the specified function and arguments. 93/// 94GenericValue JIT::runFunction(Function *F, 95 const std::vector<GenericValue> &ArgValues) { 96 assert(F && "Function *F was null at entry to run()"); 97 98 void *FPtr = getPointerToFunction(F); 99 assert(FPtr && "Pointer to fn's code was null after getPointerToFunction"); 100 const FunctionType *FTy = F->getFunctionType(); 101 const Type *RetTy = FTy->getReturnType(); 102 103 assert((FTy->getNumParams() <= ArgValues.size() || FTy->isVarArg()) && 104 "Too many arguments passed into function!"); 105 assert(FTy->getNumParams() == ArgValues.size() && 106 "This doesn't support passing arguments through varargs (yet)!"); 107 108 // Handle some common cases first. These cases correspond to common `main' 109 // prototypes. 110 if (RetTy == Type::Int32Ty || RetTy == Type::VoidTy) { 111 switch (ArgValues.size()) { 112 case 3: 113 if (FTy->getParamType(0) == Type::Int32Ty && 114 isa<PointerType>(FTy->getParamType(1)) && 115 isa<PointerType>(FTy->getParamType(2))) { 116 int (*PF)(int, char **, const char **) = 117 (int(*)(int, char **, const char **))(intptr_t)FPtr; 118 119 // Call the function. 120 GenericValue rv; 121 rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue(), 122 (char **)GVTOP(ArgValues[1]), 123 (const char **)GVTOP(ArgValues[2]))); 124 return rv; 125 } 126 break; 127 case 2: 128 if (FTy->getParamType(0) == Type::Int32Ty && 129 isa<PointerType>(FTy->getParamType(1))) { 130 int (*PF)(int, char **) = (int(*)(int, char **))(intptr_t)FPtr; 131 132 // Call the function. 133 GenericValue rv; 134 rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue(), 135 (char **)GVTOP(ArgValues[1]))); 136 return rv; 137 } 138 break; 139 case 1: 140 if (FTy->getNumParams() == 1 && 141 FTy->getParamType(0) == Type::Int32Ty) { 142 GenericValue rv; 143 int (*PF)(int) = (int(*)(int))(intptr_t)FPtr; 144 rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue())); 145 return rv; 146 } 147 break; 148 } 149 } 150 151 // Handle cases where no arguments are passed first. 152 if (ArgValues.empty()) { 153 GenericValue rv; 154 switch (RetTy->getTypeID()) { 155 default: assert(0 && "Unknown return type for function call!"); 156 case Type::IntegerTyID: { 157 unsigned BitWidth = cast<IntegerType>(RetTy)->getBitWidth(); 158 if (BitWidth == 1) 159 rv.IntVal = APInt(BitWidth, ((bool(*)())(intptr_t)FPtr)()); 160 else if (BitWidth <= 8) 161 rv.IntVal = APInt(BitWidth, ((char(*)())(intptr_t)FPtr)()); 162 else if (BitWidth <= 16) 163 rv.IntVal = APInt(BitWidth, ((short(*)())(intptr_t)FPtr)()); 164 else if (BitWidth <= 32) 165 rv.IntVal = APInt(BitWidth, ((int(*)())(intptr_t)FPtr)()); 166 else if (BitWidth <= 64) 167 rv.IntVal = APInt(BitWidth, ((int64_t(*)())(intptr_t)FPtr)()); 168 else 169 assert(0 && "Integer types > 64 bits not supported"); 170 return rv; 171 } 172 case Type::VoidTyID: 173 rv.IntVal = APInt(32, ((int(*)())(intptr_t)FPtr)()); 174 return rv; 175 case Type::FloatTyID: 176 rv.FloatVal = ((float(*)())(intptr_t)FPtr)(); 177 return rv; 178 case Type::DoubleTyID: 179 rv.DoubleVal = ((double(*)())(intptr_t)FPtr)(); 180 return rv; 181 case Type::X86_FP80TyID: 182 case Type::FP128TyID: 183 case Type::PPC_FP128TyID: 184 assert(0 && "long double not supported yet"); 185 return rv; 186 case Type::PointerTyID: 187 return PTOGV(((void*(*)())(intptr_t)FPtr)()); 188 } 189 } 190 191 // Okay, this is not one of our quick and easy cases. Because we don't have a 192 // full FFI, we have to codegen a nullary stub function that just calls the 193 // function we are interested in, passing in constants for all of the 194 // arguments. Make this function and return. 195 196 // First, create the function. 197 FunctionType *STy=FunctionType::get(RetTy, std::vector<const Type*>(), false); 198 Function *Stub = new Function(STy, Function::InternalLinkage, "", 199 F->getParent()); 200 201 // Insert a basic block. 202 BasicBlock *StubBB = new BasicBlock("", Stub); 203 204 // Convert all of the GenericValue arguments over to constants. Note that we 205 // currently don't support varargs. 206 SmallVector<Value*, 8> Args; 207 for (unsigned i = 0, e = ArgValues.size(); i != e; ++i) { 208 Constant *C = 0; 209 const Type *ArgTy = FTy->getParamType(i); 210 const GenericValue &AV = ArgValues[i]; 211 switch (ArgTy->getTypeID()) { 212 default: assert(0 && "Unknown argument type for function call!"); 213 case Type::IntegerTyID: C = ConstantInt::get(AV.IntVal); break; 214 case Type::FloatTyID: C = ConstantFP ::get(ArgTy, APFloat(AV.FloatVal)); 215 break; 216 case Type::DoubleTyID: C = ConstantFP ::get(ArgTy, APFloat(AV.DoubleVal)); 217 break; 218 case Type::PPC_FP128TyID: 219 case Type::X86_FP80TyID: 220 case Type::FP128TyID: C = ConstantFP ::get(ArgTy, APFloat(AV.IntVal)); 221 break; 222 case Type::PointerTyID: 223 void *ArgPtr = GVTOP(AV); 224 if (sizeof(void*) == 4) { 225 C = ConstantInt::get(Type::Int32Ty, (int)(intptr_t)ArgPtr); 226 } else { 227 C = ConstantInt::get(Type::Int64Ty, (intptr_t)ArgPtr); 228 } 229 C = ConstantExpr::getIntToPtr(C, ArgTy); // Cast the integer to pointer 230 break; 231 } 232 Args.push_back(C); 233 } 234 235 CallInst *TheCall = new CallInst(F, Args.begin(), Args.end(), "", StubBB); 236 TheCall->setTailCall(); 237 if (TheCall->getType() != Type::VoidTy) 238 new ReturnInst(TheCall, StubBB); // Return result of the call. 239 else 240 new ReturnInst(StubBB); // Just return void. 241 242 // Finally, return the value returned by our nullary stub function. 243 return runFunction(Stub, std::vector<GenericValue>()); 244} 245 246/// runJITOnFunction - Run the FunctionPassManager full of 247/// just-in-time compilation passes on F, hopefully filling in 248/// GlobalAddress[F] with the address of F's machine code. 249/// 250void JIT::runJITOnFunction(Function *F) { 251 static bool isAlreadyCodeGenerating = false; 252 253 MutexGuard locked(lock); 254 assert(!isAlreadyCodeGenerating && "Error: Recursive compilation detected!"); 255 256 // JIT the function 257 isAlreadyCodeGenerating = true; 258 jitstate.getPM(locked).run(*F); 259 isAlreadyCodeGenerating = false; 260 261 // If the function referred to a global variable that had not yet been 262 // emitted, it allocates memory for the global, but doesn't emit it yet. Emit 263 // all of these globals now. 264 while (!jitstate.getPendingGlobals(locked).empty()) { 265 const GlobalVariable *GV = jitstate.getPendingGlobals(locked).back(); 266 jitstate.getPendingGlobals(locked).pop_back(); 267 EmitGlobalVariable(GV); 268 } 269} 270 271/// getPointerToFunction - This method is used to get the address of the 272/// specified function, compiling it if neccesary. 273/// 274void *JIT::getPointerToFunction(Function *F) { 275 MutexGuard locked(lock); 276 277 if (void *Addr = getPointerToGlobalIfAvailable(F)) 278 return Addr; // Check if function already code gen'd 279 280 // Make sure we read in the function if it exists in this Module. 281 if (F->hasNotBeenReadFromBitcode()) { 282 // Determine the module provider this function is provided by. 283 Module *M = F->getParent(); 284 ModuleProvider *MP = 0; 285 for (unsigned i = 0, e = Modules.size(); i != e; ++i) { 286 if (Modules[i]->getModule() == M) { 287 MP = Modules[i]; 288 break; 289 } 290 } 291 assert(MP && "Function isn't in a module we know about!"); 292 293 std::string ErrorMsg; 294 if (MP->materializeFunction(F, &ErrorMsg)) { 295 cerr << "Error reading function '" << F->getName() 296 << "' from bitcode file: " << ErrorMsg << "\n"; 297 abort(); 298 } 299 } 300 301 if (F->isDeclaration()) { 302 void *Addr = getPointerToNamedFunction(F->getName()); 303 addGlobalMapping(F, Addr); 304 return Addr; 305 } 306 307 runJITOnFunction(F); 308 309 void *Addr = getPointerToGlobalIfAvailable(F); 310 assert(Addr && "Code generation didn't add function to GlobalAddress table!"); 311 return Addr; 312} 313 314/// getOrEmitGlobalVariable - Return the address of the specified global 315/// variable, possibly emitting it to memory if needed. This is used by the 316/// Emitter. 317void *JIT::getOrEmitGlobalVariable(const GlobalVariable *GV) { 318 MutexGuard locked(lock); 319 320 void *Ptr = getPointerToGlobalIfAvailable(GV); 321 if (Ptr) return Ptr; 322 323 // If the global is external, just remember the address. 324 if (GV->isDeclaration()) { 325#if HAVE___DSO_HANDLE 326 if (GV->getName() == "__dso_handle") 327 return (void*)&__dso_handle; 328#endif 329 Ptr = sys::DynamicLibrary::SearchForAddressOfSymbol(GV->getName().c_str()); 330 if (Ptr == 0) { 331 cerr << "Could not resolve external global address: " 332 << GV->getName() << "\n"; 333 abort(); 334 } 335 } else { 336 // If the global hasn't been emitted to memory yet, allocate space. We will 337 // actually initialize the global after current function has finished 338 // compilation. 339 const Type *GlobalType = GV->getType()->getElementType(); 340 size_t S = getTargetData()->getABITypeSize(GlobalType); 341 size_t A = getTargetData()->getPrefTypeAlignment(GlobalType); 342 if (A <= 8) { 343 Ptr = malloc(S); 344 } else { 345 // Allocate S+A bytes of memory, then use an aligned pointer within that 346 // space. 347 Ptr = malloc(S+A); 348 unsigned MisAligned = ((intptr_t)Ptr & (A-1)); 349 Ptr = (char*)Ptr + (MisAligned ? (A-MisAligned) : 0); 350 } 351 jitstate.getPendingGlobals(locked).push_back(GV); 352 } 353 addGlobalMapping(GV, Ptr); 354 return Ptr; 355} 356 357 358/// recompileAndRelinkFunction - This method is used to force a function 359/// which has already been compiled, to be compiled again, possibly 360/// after it has been modified. Then the entry to the old copy is overwritten 361/// with a branch to the new copy. If there was no old copy, this acts 362/// just like JIT::getPointerToFunction(). 363/// 364void *JIT::recompileAndRelinkFunction(Function *F) { 365 void *OldAddr = getPointerToGlobalIfAvailable(F); 366 367 // If it's not already compiled there is no reason to patch it up. 368 if (OldAddr == 0) { return getPointerToFunction(F); } 369 370 // Delete the old function mapping. 371 addGlobalMapping(F, 0); 372 373 // Recodegen the function 374 runJITOnFunction(F); 375 376 // Update state, forward the old function to the new function. 377 void *Addr = getPointerToGlobalIfAvailable(F); 378 assert(Addr && "Code generation didn't add function to GlobalAddress table!"); 379 TJI.replaceMachineCodeForFunction(OldAddr, Addr); 380 return Addr; 381} 382 383