llvm/lib/ExecutionEngine/JIT/JIT.cpp
Dan Gohman a3ac0c105d Reapply r57340. VMKit does not presently rely on materializeFunction
being called with the lock released, and this fixes a race condition
in the JIT as used by lli.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@64997 91177308-0d34-0410-b5e6-96231b3b80d8
2009-02-19 02:40:15 +00:00

699 lines
23 KiB
C++

//===-- JIT.cpp - LLVM Just in Time Compiler ------------------------------===//
//
// 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 "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/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/TargetJITInfo.h"
#include "llvm/Config/config.h"
using namespace llvm;
#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 <AvailabilityMacros.h>
# 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
#if HAVE___DSO_HANDLE
extern void *__dso_handle __attribute__ ((__visibility__ ("hidden")));
#endif
namespace {
static struct RegisterJIT {
RegisterJIT() { JIT::Register(); }
} JITRegistrator;
}
namespace llvm {
void LinkInJIT() {
}
}
#if defined(__GNUC__) && !defined(__ARM__EABI__)
// libgcc defines the __register_frame function to dynamically register new
// dwarf frames for exception handling. This functionality is not portable
// across compilers and is only provided by GCC. We use the __register_frame
// function here so that code generated by the JIT cooperates with the unwinding
// runtime of libgcc. When JITting with exception handling enable, LLVM
// generates dwarf frames and registers it to libgcc with __register_frame.
//
// The __register_frame function works with Linux.
//
// Unfortunately, this functionality seems to be in libgcc after the unwinding
// library of libgcc for darwin was written. The code for darwin overwrites the
// value updated by __register_frame with a value fetched with "keymgr".
// "keymgr" is an obsolete functionality, which should be rewritten some day.
// In the meantime, since "keymgr" is on all libgccs shipped with apple-gcc, we
// need a workaround in LLVM which uses the "keymgr" to dynamically modify the
// values of an opaque key, used by libgcc to find dwarf tables.
extern "C" void __register_frame(void*);
#if defined(__APPLE__)
namespace {
// LibgccObject - This is the structure defined in libgcc. There is no #include
// provided for this structure, so we also define it here. libgcc calls it
// "struct object". The structure is undocumented in libgcc.
struct LibgccObject {
void *unused1;
void *unused2;
void *unused3;
/// frame - Pointer to the exception table.
void *frame;
/// encoding - The encoding of the object?
union {
struct {
unsigned long sorted : 1;
unsigned long from_array : 1;
unsigned long mixed_encoding : 1;
unsigned long encoding : 8;
unsigned long count : 21;
} b;
size_t i;
} encoding;
/// fde_end - libgcc defines this field only if some macro is defined. We
/// include this field even if it may not there, to make libgcc happy.
char *fde_end;
/// next - At least we know it's a chained list!
struct LibgccObject *next;
};
// "kemgr" stuff. Apparently, all frame tables are stored there.
extern "C" void _keymgr_set_and_unlock_processwide_ptr(int, void *);
extern "C" void *_keymgr_get_and_lock_processwide_ptr(int);
#define KEYMGR_GCC3_DW2_OBJ_LIST 302 /* Dwarf2 object list */
/// LibgccObjectInfo - libgcc defines this struct as km_object_info. It
/// probably contains all dwarf tables that are loaded.
struct LibgccObjectInfo {
/// seenObjects - LibgccObjects already parsed by the unwinding runtime.
///
struct LibgccObject* seenObjects;
/// unseenObjects - LibgccObjects not parsed yet by the unwinding runtime.
///
struct LibgccObject* unseenObjects;
unsigned unused[2];
};
// for DW_EH_PE_omit
#include "llvm/Support/Dwarf.h"
/// darwin_register_frame - Since __register_frame does not work with darwin's
/// libgcc,we provide our own function, which "tricks" libgcc by modifying the
/// "Dwarf2 object list" key.
void DarwinRegisterFrame(void* FrameBegin) {
// Get the key.
struct LibgccObjectInfo* LOI = (struct LibgccObjectInfo*)
_keymgr_get_and_lock_processwide_ptr(KEYMGR_GCC3_DW2_OBJ_LIST);
// Allocate a new LibgccObject to represent this frame. Deallocation of this
// object may be impossible: since darwin code in libgcc was written after
// the ability to dynamically register frames, things may crash if we
// deallocate it.
struct LibgccObject* ob = (struct LibgccObject*)
malloc(sizeof(struct LibgccObject));
// Do like libgcc for the values of the field.
ob->unused1 = (void *)-1;
ob->unused2 = 0;
ob->unused3 = 0;
ob->frame = FrameBegin;
ob->encoding.i = 0;
ob->encoding.b.encoding = llvm::dwarf::DW_EH_PE_omit;
// Put the info on both places, as libgcc uses the first or the the second
// field. Note that we rely on having two pointers here. If fde_end was a
// char, things would get complicated.
ob->fde_end = (char*)LOI->unseenObjects;
ob->next = LOI->unseenObjects;
// Update the key's unseenObjects list.
LOI->unseenObjects = ob;
// Finally update the "key". Apparently, libgcc requires it.
_keymgr_set_and_unlock_processwide_ptr(KEYMGR_GCC3_DW2_OBJ_LIST,
LOI);
}
}
#endif // __APPLE__
#endif // __GNUC__
/// 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,
bool Fast) {
ExecutionEngine *EE = JIT::createJIT(MP, ErrorStr, JMM, Fast);
if (!EE) return 0;
// 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;
}
JIT::JIT(ModuleProvider *MP, TargetMachine &tm, TargetJITInfo &tji,
JITMemoryManager *JMM, bool Fast)
: ExecutionEngine(MP), TM(tm), TJI(tji) {
setTargetData(TM.getTargetData());
jitstate = new JITState(MP);
// Initialize MCE
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, Fast)) {
cerr << "Target does not support machine code emission!\n";
abort();
}
// Register routine for informing unwinding runtime about new EH frames
#if defined(__GNUC__) && !defined(__ARM_EABI__)
#if defined(__APPLE__)
struct LibgccObjectInfo* LOI = (struct LibgccObjectInfo*)
_keymgr_get_and_lock_processwide_ptr(KEYMGR_GCC3_DW2_OBJ_LIST);
// The key is created on demand, and libgcc creates it the first time an
// exception occurs. Since we need the key to register frames, we create
// it now.
if (!LOI) {
LOI = (LibgccObjectInfo*)malloc(sizeof(struct LibgccObjectInfo));
_keymgr_set_and_unlock_processwide_ptr(KEYMGR_GCC3_DW2_OBJ_LIST,
LOI);
}
InstallExceptionTableRegister(DarwinRegisterFrame);
#else
InstallExceptionTableRegister(__register_frame);
#endif // __APPLE__
#endif // __GNUC__
// Initialize passes.
PM.doInitialization();
}
JIT::~JIT() {
delete jitstate;
delete MCE;
delete &TM;
}
/// addModuleProvider - Add a new ModuleProvider to the JIT. If we previously
/// removed the last ModuleProvider, we need re-initialize jitstate with a valid
/// ModuleProvider.
void JIT::addModuleProvider(ModuleProvider *MP) {
MutexGuard locked(lock);
if (Modules.empty()) {
assert(!jitstate && "jitstate should be NULL if Modules vector is empty!");
jitstate = new JITState(MP);
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();
}
// Initialize passes.
PM.doInitialization();
}
ExecutionEngine::addModuleProvider(MP);
}
/// removeModuleProvider - If we are removing the last ModuleProvider,
/// invalidate the jitstate since the PassManager it contains references a
/// released ModuleProvider.
Module *JIT::removeModuleProvider(ModuleProvider *MP, std::string *E) {
Module *result = ExecutionEngine::removeModuleProvider(MP, E);
MutexGuard locked(lock);
if (jitstate->getMP() == MP) {
delete jitstate;
jitstate = 0;
}
if (!jitstate && !Modules.empty()) {
jitstate = new JITState(Modules[0]);
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();
}
// Initialize passes.
PM.doInitialization();
}
return result;
}
/// deleteModuleProvider - Remove a ModuleProvider from the list of modules,
/// and deletes the ModuleProvider and owned Module. Avoids materializing
/// the underlying module.
void JIT::deleteModuleProvider(ModuleProvider *MP, std::string *E) {
ExecutionEngine::deleteModuleProvider(MP, E);
MutexGuard locked(lock);
if (jitstate->getMP() == MP) {
delete jitstate;
jitstate = 0;
}
if (!jitstate && !Modules.empty()) {
jitstate = new JITState(Modules[0]);
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();
}
// Initialize passes.
PM.doInitialization();
}
}
/// run - Start execution with the specified function and arguments.
///
GenericValue JIT::runFunction(Function *F,
const std::vector<GenericValue> &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<PointerType>(FTy->getParamType(1)) &&
isa<PointerType>(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<PointerType>(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<IntegerType>(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<const Type*>(), 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<Value*, 8> 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->setCallingConv(F->getCallingConv());
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<GenericValue>());
}
/// 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.
///
void JIT::runJITOnFunction(Function *F) {
MutexGuard locked(lock);
runJITOnFunctionUnlocked(F, locked);
}
void JIT::runJITOnFunctionUnlocked(Function *F, const MutexGuard &locked) {
static bool isAlreadyCodeGenerating = false;
assert(!isAlreadyCodeGenerating && "Error: Recursive compilation detected!");
// JIT the function
isAlreadyCodeGenerating = true;
jitstate->getPM(locked).run(*F);
isAlreadyCodeGenerating = false;
// If the function referred to another function that had not yet been
// read from bitcode, but we are jitting non-lazily, emit it now.
while (!jitstate->getPendingFunctions(locked).empty()) {
Function *PF = jitstate->getPendingFunctions(locked).back();
jitstate->getPendingFunctions(locked).pop_back();
// JIT the function
isAlreadyCodeGenerating = true;
jitstate->getPM(locked).run(*PF);
isAlreadyCodeGenerating = false;
// Now that the function has been jitted, ask the JITEmitter to rewrite
// the stub with real address of the function.
updateFunctionStub(PF);
}
// If the JIT is configured to emit info so that dlsym can be used to
// rewrite stubs to external globals, do so now.
if (areDlsymStubsEnabled() && isLazyCompilationDisabled())
updateDlsymStubTable();
}
/// getPointerToFunction - This method is used to get the address of the
/// specified function, compiling it if neccesary.
///
void *JIT::getPointerToFunction(Function *F) {
if (void *Addr = getPointerToGlobalIfAvailable(F))
return Addr; // Check if function already code gen'd
MutexGuard locked(lock);
// 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();
}
// Now retry to get the address.
if (void *Addr = getPointerToGlobalIfAvailable(F))
return Addr;
}
if (F->isDeclaration()) {
bool AbortOnFailure = F->getLinkage() != GlobalValue::ExternalWeakLinkage;
void *Addr = getPointerToNamedFunction(F->getName(), AbortOnFailure);
addGlobalMapping(F, Addr);
return Addr;
}
runJITOnFunctionUnlocked(F, locked);
void *Addr = getPointerToGlobalIfAvailable(F);
assert(Addr && "Code generation didn't add function to GlobalAddress table!");
return Addr;
}
/// 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);
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();
addGlobalMapping(GV, Ptr);
}
} else {
// GlobalVariable's which are not "constant" will cause trouble in a server
// situation. It's returned in the same block of memory as code which may
// not be writable.
if (isGVCompilationDisabled() && !GV->isConstant()) {
cerr << "Compilation of non-internal GlobalValue is disabled!\n";
abort();
}
// If the global hasn't been emitted to memory yet, allocate space and
// emit it into memory. It goes in the same array as the generated
// code, jump tables, etc.
const Type *GlobalType = GV->getType()->getElementType();
size_t S = getTargetData()->getTypePaddedSize(GlobalType);
size_t A = getTargetData()->getPreferredAlignment(GV);
if (GV->isThreadLocal()) {
MutexGuard locked(lock);
Ptr = TJI.allocateThreadLocalMemory(S);
} else if (TJI.allocateSeparateGVMemory()) {
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);
}
} else {
Ptr = MCE->allocateSpace(S, A);
}
addGlobalMapping(GV, Ptr);
EmitGlobalVariable(GV);
}
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;
}
/// getMemoryForGV - This method abstracts memory allocation of global
/// variable so that the JIT can allocate thread local variables depending
/// on the target.
///
char* JIT::getMemoryForGV(const GlobalVariable* GV) {
const Type *ElTy = GV->getType()->getElementType();
size_t GVSize = (size_t)getTargetData()->getTypePaddedSize(ElTy);
if (GV->isThreadLocal()) {
MutexGuard locked(lock);
return TJI.allocateThreadLocalMemory(GVSize);
} else {
return new char[GVSize];
}
}
void JIT::addPendingFunction(Function *F) {
MutexGuard locked(lock);
jitstate->getPendingFunctions(locked).push_back(F);
}