X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;ds=sidebyside;f=lib%2FExecutionEngine%2FJIT%2FJIT.cpp;h=4b888ef018ce42c121557d1d1985d8b00eeb3d22;hb=b1dbcd886a4b5597a839f299054b78b33fb2d6df;hp=7f4877b9cb264ace3e1e5d823e57cd55caca2a68;hpb=05a1a306bcfba2b533dc1210d24924e5f3e9ed0e;p=oota-llvm.git diff --git a/lib/ExecutionEngine/JIT/JIT.cpp b/lib/ExecutionEngine/JIT/JIT.cpp index 7f4877b9cb2..4b888ef018c 100644 --- a/lib/ExecutionEngine/JIT/JIT.cpp +++ b/lib/ExecutionEngine/JIT/JIT.cpp @@ -1,138 +1,421 @@ //===-- 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. +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This tool implements a just-in-time compiler for LLVM, allowing direct +// execution of LLVM bitcode in an efficient manner. // //===----------------------------------------------------------------------===// -#include "VM.h" +#include "JIT.h" +#include "llvm/Constants.h" +#include "llvm/DerivedTypes.h" +#include "llvm/Function.h" +#include "llvm/GlobalVariable.h" +#include "llvm/Instructions.h" +#include "llvm/ModuleProvider.h" +#include "llvm/CodeGen/MachineCodeEmitter.h" +#include "llvm/CodeGen/MachineFunction.h" +#include "llvm/ExecutionEngine/GenericValue.h" +#include "llvm/Support/MutexGuard.h" +#include "llvm/System/DynamicLibrary.h" +#include "llvm/Target/TargetData.h" #include "llvm/Target/TargetMachine.h" -#include "llvm/Target/TargetMachineImpls.h" -#include "llvm/Module.h" -#include "Support/CommandLine.h" +#include "llvm/Target/TargetJITInfo.h" -// FIXME: REMOVE THIS -#include "llvm/PassManager.h" +#include "llvm/Config/config.h" -#if !defined(ENABLE_X86_JIT) && !defined(ENABLE_SPARC_JIT) -#define NO_JITS_ENABLED -#endif +using namespace llvm; -namespace { - enum ArchName { x86, Sparc }; - -#ifndef NO_JITS_ENABLED - cl::opt - Arch("march", cl::desc("Architecture to JIT to:"), cl::Prefix, - cl::values( -#ifdef ENABLE_X86_JIT - clEnumVal(x86, " IA-32 (Pentium and above)"), -#endif -#ifdef ENABLE_SPARC_JIT - clEnumValN(Sparc, "sparc", " Sparc-V9"), +#ifdef __APPLE__ +// Apple gcc defaults to -fuse-cxa-atexit (i.e. calls __cxa_atexit instead +// of atexit). It passes the address of linker generated symbol __dso_handle +// to the function. +// This configuration change happened at version 5330. +# include +# if defined(MAC_OS_X_VERSION_10_4) && \ + ((MAC_OS_X_VERSION_MIN_REQUIRED > MAC_OS_X_VERSION_10_4) || \ + (MAC_OS_X_VERSION_MIN_REQUIRED == MAC_OS_X_VERSION_10_4 && \ + __APPLE_CC__ >= 5330)) +# ifndef HAVE___DSO_HANDLE +# define HAVE___DSO_HANDLE 1 +# endif +# endif #endif - 0), -#if defined(ENABLE_X86_JIT) - cl::init(x86) -#elif defined(ENABLE_SPARC_JIT) - cl::init(Sparc) + +#if HAVE___DSO_HANDLE +extern void *__dso_handle __attribute__ ((__visibility__ ("hidden"))); #endif - ); -#endif /* NO_JITS_ENABLED */ + +namespace { + +static struct RegisterJIT { + RegisterJIT() { JIT::Register(); } +} JITRegistrator; + } -/// 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) { - - TargetMachine* (*TargetMachineAllocator)(unsigned) = 0; +namespace llvm { + void LinkInJIT() { + } +} - // Allow a command-line switch to override what *should* be the default target - // machine for this platform. This allows for debugging a Sparc JIT on X86 -- - // our X86 machines are much faster at recompiling LLVM and linking LLI. -#ifdef NO_JITS_ENABLED - return 0; +#if defined (__GNUC__) +extern "C" void __register_frame(void*); #endif - switch (Arch) { -#ifdef ENABLE_X86_JIT - case x86: - TargetMachineAllocator = allocateX86TargetMachine; - break; -#endif -#ifdef ENABLE_SPARC_JIT - case Sparc: - TargetMachineAllocator = allocateSparcTargetMachine; - break; +/// createJIT - This is the factory method for creating a JIT for the current +/// machine, it does not fall back to the interpreter. This takes ownership +/// of the module provider. +ExecutionEngine *ExecutionEngine::createJIT(ModuleProvider *MP, + std::string *ErrorStr, + JITMemoryManager *JMM) { + ExecutionEngine *EE = JIT::createJIT(MP, ErrorStr, JMM); + if (!EE) return 0; + + // Register routine for informing unwinding runtime about new EH frames +#if defined(__GNUC__) + EE->InstallExceptionTableRegister(__register_frame); #endif - default: - assert(0 && "-march flag not supported on this host!"); - } - // Allocate a target... - TargetMachine *Target = (*TargetMachineAllocator)(Config); - assert(Target && "Could not allocate target machine!"); - - // Create the virtual machine object... - return new VM(M, Target); + // Make sure we can resolve symbols in the program as well. The zero arg + // to the function tells DynamicLibrary to load the program, not a library. + sys::DynamicLibrary::LoadLibraryPermanently(0, ErrorStr); + return EE; } -VM::VM(Module *M, TargetMachine *tm) : ExecutionEngine(M), TM(*tm) { +JIT::JIT(ModuleProvider *MP, TargetMachine &tm, TargetJITInfo &tji, + JITMemoryManager *JMM) + : ExecutionEngine(MP), TM(tm), TJI(tji), jitstate(MP) { setTargetData(TM.getTargetData()); // Initialize MCE - MCE = createEmitter(*this); - - setupPassManager(); - -#ifdef ENABLE_SPARC_JIT - // THIS GOES BEYOND UGLY HACKS - if (TM.getName() == "UltraSparc-Native") { - extern Pass *createPreSelectionPass(TargetMachine &TM); - PassManager PM; - // Specialize LLVM code for this target machine and then - // run basic dataflow optimizations on LLVM code. - PM.add(createPreSelectionPass(TM)); - PM.run(*M); + MCE = createEmitter(*this, JMM); + + // Add target data + MutexGuard locked(lock); + FunctionPassManager &PM = jitstate.getPM(locked); + PM.add(new TargetData(*TM.getTargetData())); + + // Turn the machine code intermediate representation into bytes in memory that + // may be executed. + if (TM.addPassesToEmitMachineCode(PM, *MCE, false /*fast*/)) { + cerr << "Target does not support machine code emission!\n"; + abort(); } -#endif + + // Initialize passes. + PM.doInitialization(); +} - emitGlobals(); +JIT::~JIT() { + delete MCE; + delete &TM; } -/// VM::run - This method begins the execution of a program beginning at the -/// specified function name. The function is called with the specified -/// arguments and array of environment variables (a la main()). +/// run - Start execution with the specified function and arguments. /// -/// Inputs: -/// FnName - The name of the function as a C++ string. -/// Args - A vector of C++ strings containing the arguments. -/// envp - An array of C strings containing the environment. +GenericValue JIT::runFunction(Function *F, + const std::vector &ArgValues) { + assert(F && "Function *F was null at entry to run()"); + + void *FPtr = getPointerToFunction(F); + assert(FPtr && "Pointer to fn's code was null after getPointerToFunction"); + const FunctionType *FTy = F->getFunctionType(); + const Type *RetTy = FTy->getReturnType(); + + assert((FTy->getNumParams() <= ArgValues.size() || FTy->isVarArg()) && + "Too many arguments passed into function!"); + assert(FTy->getNumParams() == ArgValues.size() && + "This doesn't support passing arguments through varargs (yet)!"); + + // Handle some common cases first. These cases correspond to common `main' + // prototypes. + if (RetTy == Type::Int32Ty || RetTy == Type::VoidTy) { + switch (ArgValues.size()) { + case 3: + if (FTy->getParamType(0) == Type::Int32Ty && + isa(FTy->getParamType(1)) && + isa(FTy->getParamType(2))) { + int (*PF)(int, char **, const char **) = + (int(*)(int, char **, const char **))(intptr_t)FPtr; + + // Call the function. + GenericValue rv; + rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue(), + (char **)GVTOP(ArgValues[1]), + (const char **)GVTOP(ArgValues[2]))); + return rv; + } + break; + case 2: + if (FTy->getParamType(0) == Type::Int32Ty && + isa(FTy->getParamType(1))) { + int (*PF)(int, char **) = (int(*)(int, char **))(intptr_t)FPtr; + + // Call the function. + GenericValue rv; + rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue(), + (char **)GVTOP(ArgValues[1]))); + return rv; + } + break; + case 1: + if (FTy->getNumParams() == 1 && + FTy->getParamType(0) == Type::Int32Ty) { + GenericValue rv; + int (*PF)(int) = (int(*)(int))(intptr_t)FPtr; + rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue())); + return rv; + } + break; + } + } + + // Handle cases where no arguments are passed first. + if (ArgValues.empty()) { + GenericValue rv; + switch (RetTy->getTypeID()) { + default: assert(0 && "Unknown return type for function call!"); + case Type::IntegerTyID: { + unsigned BitWidth = cast(RetTy)->getBitWidth(); + if (BitWidth == 1) + rv.IntVal = APInt(BitWidth, ((bool(*)())(intptr_t)FPtr)()); + else if (BitWidth <= 8) + rv.IntVal = APInt(BitWidth, ((char(*)())(intptr_t)FPtr)()); + else if (BitWidth <= 16) + rv.IntVal = APInt(BitWidth, ((short(*)())(intptr_t)FPtr)()); + else if (BitWidth <= 32) + rv.IntVal = APInt(BitWidth, ((int(*)())(intptr_t)FPtr)()); + else if (BitWidth <= 64) + rv.IntVal = APInt(BitWidth, ((int64_t(*)())(intptr_t)FPtr)()); + else + assert(0 && "Integer types > 64 bits not supported"); + return rv; + } + case Type::VoidTyID: + rv.IntVal = APInt(32, ((int(*)())(intptr_t)FPtr)()); + return rv; + case Type::FloatTyID: + rv.FloatVal = ((float(*)())(intptr_t)FPtr)(); + return rv; + case Type::DoubleTyID: + rv.DoubleVal = ((double(*)())(intptr_t)FPtr)(); + return rv; + case Type::X86_FP80TyID: + case Type::FP128TyID: + case Type::PPC_FP128TyID: + assert(0 && "long double not supported yet"); + return rv; + case Type::PointerTyID: + return PTOGV(((void*(*)())(intptr_t)FPtr)()); + } + } + + // Okay, this is not one of our quick and easy cases. Because we don't have a + // full FFI, we have to codegen a nullary stub function that just calls the + // function we are interested in, passing in constants for all of the + // arguments. Make this function and return. + + // First, create the function. + FunctionType *STy=FunctionType::get(RetTy, std::vector(), false); + Function *Stub = Function::Create(STy, Function::InternalLinkage, "", + F->getParent()); + + // Insert a basic block. + BasicBlock *StubBB = BasicBlock::Create("", Stub); + + // Convert all of the GenericValue arguments over to constants. Note that we + // currently don't support varargs. + SmallVector Args; + for (unsigned i = 0, e = ArgValues.size(); i != e; ++i) { + Constant *C = 0; + const Type *ArgTy = FTy->getParamType(i); + const GenericValue &AV = ArgValues[i]; + switch (ArgTy->getTypeID()) { + default: assert(0 && "Unknown argument type for function call!"); + case Type::IntegerTyID: + C = ConstantInt::get(AV.IntVal); + break; + case Type::FloatTyID: + C = ConstantFP::get(APFloat(AV.FloatVal)); + break; + case Type::DoubleTyID: + C = ConstantFP::get(APFloat(AV.DoubleVal)); + break; + case Type::PPC_FP128TyID: + case Type::X86_FP80TyID: + case Type::FP128TyID: + C = ConstantFP::get(APFloat(AV.IntVal)); + break; + case Type::PointerTyID: + void *ArgPtr = GVTOP(AV); + if (sizeof(void*) == 4) + C = ConstantInt::get(Type::Int32Ty, (int)(intptr_t)ArgPtr); + else + C = ConstantInt::get(Type::Int64Ty, (intptr_t)ArgPtr); + C = ConstantExpr::getIntToPtr(C, ArgTy); // Cast the integer to pointer + break; + } + Args.push_back(C); + } + + CallInst *TheCall = CallInst::Create(F, Args.begin(), Args.end(), + "", StubBB); + TheCall->setTailCall(); + if (TheCall->getType() != Type::VoidTy) + ReturnInst::Create(TheCall, StubBB); // Return result of the call. + else + ReturnInst::Create(StubBB); // Just return void. + + // Finally, return the value returned by our nullary stub function. + return runFunction(Stub, std::vector()); +} + +/// runJITOnFunction - Run the FunctionPassManager full of +/// just-in-time compilation passes on F, hopefully filling in +/// GlobalAddress[F] with the address of F's machine code. /// -/// Return value: -/// 1 - An error occurred. -/// Otherwise, the return value from the specified function is returned. +void JIT::runJITOnFunction(Function *F) { + static bool isAlreadyCodeGenerating = false; + + MutexGuard locked(lock); + assert(!isAlreadyCodeGenerating && "Error: Recursive compilation detected!"); + + // JIT the function + isAlreadyCodeGenerating = true; + jitstate.getPM(locked).run(*F); + isAlreadyCodeGenerating = false; + + // If the function referred to a global variable that had not yet been + // emitted, it allocates memory for the global, but doesn't emit it yet. Emit + // all of these globals now. + while (!jitstate.getPendingGlobals(locked).empty()) { + const GlobalVariable *GV = jitstate.getPendingGlobals(locked).back(); + jitstate.getPendingGlobals(locked).pop_back(); + EmitGlobalVariable(GV); + } +} + +/// getPointerToFunction - This method is used to get the address of the +/// specified function, compiling it if neccesary. /// -int VM::run(const std::string &FnName, const std::vector &Args, - const char **envp) { - Function *F = getModule().getNamedFunction(FnName); - if (F == 0) { - std::cerr << "Could not find function '" << FnName << "' in module!\n"; - return 1; +void *JIT::getPointerToFunction(Function *F) { + + if (void *Addr = getPointerToGlobalIfAvailable(F)) + return Addr; // Check if function already code gen'd + + // Make sure we read in the function if it exists in this Module. + if (F->hasNotBeenReadFromBitcode()) { + // Determine the module provider this function is provided by. + Module *M = F->getParent(); + ModuleProvider *MP = 0; + for (unsigned i = 0, e = Modules.size(); i != e; ++i) { + if (Modules[i]->getModule() == M) { + MP = Modules[i]; + break; + } + } + assert(MP && "Function isn't in a module we know about!"); + + std::string ErrorMsg; + if (MP->materializeFunction(F, &ErrorMsg)) { + cerr << "Error reading function '" << F->getName() + << "' from bitcode file: " << ErrorMsg << "\n"; + abort(); + } + } + + if (void *Addr = getPointerToGlobalIfAvailable(F)) { + return Addr; } - int (*PF)(int, char**, const char**) = - (int(*)(int, char**, const char**))getPointerToFunction(F); - assert(PF != 0 && "Null pointer to function?"); + MutexGuard locked(lock); + + if (F->isDeclaration()) { + void *Addr = getPointerToNamedFunction(F->getName()); + addGlobalMapping(F, Addr); + return Addr; + } - // Build an argv vector... - char **Argv = (char**)CreateArgv(Args); + runJITOnFunction(F); + + void *Addr = getPointerToGlobalIfAvailable(F); + assert(Addr && "Code generation didn't add function to GlobalAddress table!"); + return Addr; +} - // Call the main function... - int Result = PF(Args.size(), Argv, envp); +/// getOrEmitGlobalVariable - Return the address of the specified global +/// variable, possibly emitting it to memory if needed. This is used by the +/// Emitter. +void *JIT::getOrEmitGlobalVariable(const GlobalVariable *GV) { + MutexGuard locked(lock); - // Run any atexit handlers now! - runAtExitHandlers(); - return Result; + void *Ptr = getPointerToGlobalIfAvailable(GV); + if (Ptr) return Ptr; + + // If the global is external, just remember the address. + if (GV->isDeclaration()) { +#if HAVE___DSO_HANDLE + if (GV->getName() == "__dso_handle") + return (void*)&__dso_handle; +#endif + Ptr = sys::DynamicLibrary::SearchForAddressOfSymbol(GV->getName().c_str()); + if (Ptr == 0) { + cerr << "Could not resolve external global address: " + << GV->getName() << "\n"; + abort(); + } + } else { + // If the global hasn't been emitted to memory yet, allocate space. We will + // actually initialize the global after current function has finished + // compilation. + const Type *GlobalType = GV->getType()->getElementType(); + size_t S = getTargetData()->getABITypeSize(GlobalType); + size_t A = getTargetData()->getPreferredAlignment(GV); + if (A <= 8) { + Ptr = malloc(S); + } else { + // Allocate S+A bytes of memory, then use an aligned pointer within that + // space. + Ptr = malloc(S+A); + unsigned MisAligned = ((intptr_t)Ptr & (A-1)); + Ptr = (char*)Ptr + (MisAligned ? (A-MisAligned) : 0); + } + jitstate.getPendingGlobals(locked).push_back(GV); + } + addGlobalMapping(GV, Ptr); + return Ptr; } + + +/// recompileAndRelinkFunction - This method is used to force a function +/// which has already been compiled, to be compiled again, possibly +/// after it has been modified. Then the entry to the old copy is overwritten +/// with a branch to the new copy. If there was no old copy, this acts +/// just like JIT::getPointerToFunction(). +/// +void *JIT::recompileAndRelinkFunction(Function *F) { + void *OldAddr = getPointerToGlobalIfAvailable(F); + + // If it's not already compiled there is no reason to patch it up. + if (OldAddr == 0) { return getPointerToFunction(F); } + + // Delete the old function mapping. + addGlobalMapping(F, 0); + + // Recodegen the function + runJITOnFunction(F); + + // Update state, forward the old function to the new function. + void *Addr = getPointerToGlobalIfAvailable(F); + assert(Addr && "Code generation didn't add function to GlobalAddress table!"); + TJI.replaceMachineCodeForFunction(OldAddr, Addr); + return Addr; +} +