1 //===-- JIT.cpp - LLVM Just in Time Compiler ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This tool implements a just-in-time compiler for LLVM, allowing direct
11 // execution of LLVM bitcode in an efficient manner.
13 //===----------------------------------------------------------------------===//
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/ExecutionEngine/GenericValue.h"
24 #include "llvm/Support/MutexGuard.h"
25 #include "llvm/System/DynamicLibrary.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetMachine.h"
28 #include "llvm/Target/TargetJITInfo.h"
30 #include "llvm/Config/config.h"
35 // Apple gcc defaults to -fuse-cxa-atexit (i.e. calls __cxa_atexit instead
36 // of atexit). It passes the address of linker generated symbol __dso_handle
38 // This configuration change happened at version 5330.
39 # include <AvailabilityMacros.h>
40 # if defined(MAC_OS_X_VERSION_10_4) && \
41 ((MAC_OS_X_VERSION_MIN_REQUIRED > MAC_OS_X_VERSION_10_4) || \
42 (MAC_OS_X_VERSION_MIN_REQUIRED == MAC_OS_X_VERSION_10_4 && \
43 __APPLE_CC__ >= 5330))
44 # ifndef HAVE___DSO_HANDLE
45 # define HAVE___DSO_HANDLE 1
51 extern void *__dso_handle __attribute__ ((__visibility__ ("hidden")));
56 static struct RegisterJIT {
57 RegisterJIT() { JIT::Register(); }
68 #if defined(__GNUC__) && !defined(__ARM__EABI__)
70 // libgcc defines the __register_frame function to dynamically register new
71 // dwarf frames for exception handling. This functionality is not portable
72 // across compilers and is only provided by GCC. We use the __register_frame
73 // function here so that code generated by the JIT cooperates with the unwinding
74 // runtime of libgcc. When JITting with exception handling enable, LLVM
75 // generates dwarf frames and registers it to libgcc with __register_frame.
77 // The __register_frame function works with Linux.
79 // Unfortunately, this functionality seems to be in libgcc after the unwinding
80 // library of libgcc for darwin was written. The code for darwin overwrites the
81 // value updated by __register_frame with a value fetched with "keymgr".
82 // "keymgr" is an obsolete functionality, which should be rewritten some day.
83 // In the meantime, since "keymgr" is on all libgccs shipped with apple-gcc, we
84 // need a workaround in LLVM which uses the "keymgr" to dynamically modify the
85 // values of an opaque key, used by libgcc to find dwarf tables.
87 extern "C" void __register_frame(void*);
89 #if defined(__APPLE__)
93 // LibgccObject - This is the structure defined in libgcc. There is no #include
94 // provided for this structure, so we also define it here. libgcc calls it
95 // "struct object". The structure is undocumented in libgcc.
101 /// frame - Pointer to the exception table.
104 /// encoding - The encoding of the object?
107 unsigned long sorted : 1;
108 unsigned long from_array : 1;
109 unsigned long mixed_encoding : 1;
110 unsigned long encoding : 8;
111 unsigned long count : 21;
116 /// fde_end - libgcc defines this field only if some macro is defined. We
117 /// include this field even if it may not there, to make libgcc happy.
120 /// next - At least we know it's a chained list!
121 struct LibgccObject *next;
124 // "kemgr" stuff. Apparently, all frame tables are stored there.
125 extern "C" void _keymgr_set_and_unlock_processwide_ptr(int, void *);
126 extern "C" void *_keymgr_get_and_lock_processwide_ptr(int);
127 #define KEYMGR_GCC3_DW2_OBJ_LIST 302 /* Dwarf2 object list */
129 /// LibgccObjectInfo - libgcc defines this struct as km_object_info. It
130 /// probably contains all dwarf tables that are loaded.
131 struct LibgccObjectInfo {
133 /// seenObjects - LibgccObjects already parsed by the unwinding runtime.
135 struct LibgccObject* seenObjects;
137 /// unseenObjects - LibgccObjects not parsed yet by the unwinding runtime.
139 struct LibgccObject* unseenObjects;
145 #include "llvm/Support/Dwarf.h"
147 /// darwin_register_frame - Since __register_frame does not work with darwin's
148 /// libgcc,we provide our own function, which "tricks" libgcc by modifying the
149 /// "Dwarf2 object list" key.
150 void DarwinRegisterFrame(void* FrameBegin) {
152 struct LibgccObjectInfo* LOI = (struct LibgccObjectInfo*)
153 _keymgr_get_and_lock_processwide_ptr(KEYMGR_GCC3_DW2_OBJ_LIST);
155 // Allocate a new LibgccObject to represent this frame. Deallocation of this
156 // object may be impossible: since darwin code in libgcc was written after
157 // the ability to dynamically register frames, things may crash if we
159 struct LibgccObject* ob = (struct LibgccObject*)
160 malloc(sizeof(struct LibgccObject));
162 // Do like libgcc for the values of the field.
163 ob->unused1 = (void *)-1;
166 ob->frame = FrameBegin;
168 ob->encoding.b.encoding = llvm::dwarf::DW_EH_PE_omit;
170 // Put the info on both places, as libgcc uses the first or the the second
171 // field. Note that we rely on having two pointers here. If fde_end was a
172 // char, things would get complicated.
173 ob->fde_end = (char*)LOI->unseenObjects;
174 ob->next = LOI->unseenObjects;
176 // Update the key's unseenObjects list.
177 LOI->unseenObjects = ob;
179 // Finally update the "key". Apparently, libgcc requires it.
180 _keymgr_set_and_unlock_processwide_ptr(KEYMGR_GCC3_DW2_OBJ_LIST,
189 /// createJIT - This is the factory method for creating a JIT for the current
190 /// machine, it does not fall back to the interpreter. This takes ownership
191 /// of the module provider.
192 ExecutionEngine *ExecutionEngine::createJIT(ModuleProvider *MP,
193 std::string *ErrorStr,
194 JITMemoryManager *JMM,
196 ExecutionEngine *EE = JIT::createJIT(MP, ErrorStr, JMM, Fast);
199 // Make sure we can resolve symbols in the program as well. The zero arg
200 // to the function tells DynamicLibrary to load the program, not a library.
201 sys::DynamicLibrary::LoadLibraryPermanently(0, ErrorStr);
205 JIT::JIT(ModuleProvider *MP, TargetMachine &tm, TargetJITInfo &tji,
206 JITMemoryManager *JMM, bool Fast)
207 : ExecutionEngine(MP), TM(tm), TJI(tji) {
208 setTargetData(TM.getTargetData());
210 jitstate = new JITState(MP);
213 MCE = createEmitter(*this, JMM);
216 MutexGuard locked(lock);
217 FunctionPassManager &PM = jitstate->getPM(locked);
218 PM.add(new TargetData(*TM.getTargetData()));
220 // Turn the machine code intermediate representation into bytes in memory that
222 if (TM.addPassesToEmitMachineCode(PM, *MCE, Fast)) {
223 cerr << "Target does not support machine code emission!\n";
227 // Register routine for informing unwinding runtime about new EH frames
228 #if defined(__GNUC__) && !defined(__ARM_EABI__)
229 #if defined(__APPLE__)
230 struct LibgccObjectInfo* LOI = (struct LibgccObjectInfo*)
231 _keymgr_get_and_lock_processwide_ptr(KEYMGR_GCC3_DW2_OBJ_LIST);
233 // The key is created on demand, and libgcc creates it the first time an
234 // exception occurs. Since we need the key to register frames, we create
237 LOI = (LibgccObjectInfo*)malloc(sizeof(struct LibgccObjectInfo));
238 _keymgr_set_and_unlock_processwide_ptr(KEYMGR_GCC3_DW2_OBJ_LIST,
241 InstallExceptionTableRegister(DarwinRegisterFrame);
243 InstallExceptionTableRegister(__register_frame);
247 // Initialize passes.
248 PM.doInitialization();
257 /// addModuleProvider - Add a new ModuleProvider to the JIT. If we previously
258 /// removed the last ModuleProvider, we need re-initialize jitstate with a valid
260 void JIT::addModuleProvider(ModuleProvider *MP) {
261 MutexGuard locked(lock);
263 if (Modules.empty()) {
264 assert(!jitstate && "jitstate should be NULL if Modules vector is empty!");
266 jitstate = new JITState(MP);
268 FunctionPassManager &PM = jitstate->getPM(locked);
269 PM.add(new TargetData(*TM.getTargetData()));
271 // Turn the machine code intermediate representation into bytes in memory
272 // that may be executed.
273 if (TM.addPassesToEmitMachineCode(PM, *MCE, false /*fast*/)) {
274 cerr << "Target does not support machine code emission!\n";
278 // Initialize passes.
279 PM.doInitialization();
282 ExecutionEngine::addModuleProvider(MP);
285 /// removeModuleProvider - If we are removing the last ModuleProvider,
286 /// invalidate the jitstate since the PassManager it contains references a
287 /// released ModuleProvider.
288 Module *JIT::removeModuleProvider(ModuleProvider *MP, std::string *E) {
289 Module *result = ExecutionEngine::removeModuleProvider(MP, E);
291 MutexGuard locked(lock);
292 if (Modules.empty()) {
300 /// deleteModuleProvider - Remove a ModuleProvider from the list of modules,
301 /// and deletes the ModuleProvider and owned Module. Avoids materializing
302 /// the underlying module.
303 void JIT::deleteModuleProvider(ModuleProvider *MP, std::string *E) {
304 ExecutionEngine::deleteModuleProvider(MP, E);
306 MutexGuard locked(lock);
307 if (Modules.empty()) {
313 /// run - Start execution with the specified function and arguments.
315 GenericValue JIT::runFunction(Function *F,
316 const std::vector<GenericValue> &ArgValues) {
317 assert(F && "Function *F was null at entry to run()");
319 void *FPtr = getPointerToFunction(F);
320 assert(FPtr && "Pointer to fn's code was null after getPointerToFunction");
321 const FunctionType *FTy = F->getFunctionType();
322 const Type *RetTy = FTy->getReturnType();
324 assert((FTy->getNumParams() <= ArgValues.size() || FTy->isVarArg()) &&
325 "Too many arguments passed into function!");
326 assert(FTy->getNumParams() == ArgValues.size() &&
327 "This doesn't support passing arguments through varargs (yet)!");
329 // Handle some common cases first. These cases correspond to common `main'
331 if (RetTy == Type::Int32Ty || RetTy == Type::VoidTy) {
332 switch (ArgValues.size()) {
334 if (FTy->getParamType(0) == Type::Int32Ty &&
335 isa<PointerType>(FTy->getParamType(1)) &&
336 isa<PointerType>(FTy->getParamType(2))) {
337 int (*PF)(int, char **, const char **) =
338 (int(*)(int, char **, const char **))(intptr_t)FPtr;
340 // Call the function.
342 rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue(),
343 (char **)GVTOP(ArgValues[1]),
344 (const char **)GVTOP(ArgValues[2])));
349 if (FTy->getParamType(0) == Type::Int32Ty &&
350 isa<PointerType>(FTy->getParamType(1))) {
351 int (*PF)(int, char **) = (int(*)(int, char **))(intptr_t)FPtr;
353 // Call the function.
355 rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue(),
356 (char **)GVTOP(ArgValues[1])));
361 if (FTy->getNumParams() == 1 &&
362 FTy->getParamType(0) == Type::Int32Ty) {
364 int (*PF)(int) = (int(*)(int))(intptr_t)FPtr;
365 rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue()));
372 // Handle cases where no arguments are passed first.
373 if (ArgValues.empty()) {
375 switch (RetTy->getTypeID()) {
376 default: assert(0 && "Unknown return type for function call!");
377 case Type::IntegerTyID: {
378 unsigned BitWidth = cast<IntegerType>(RetTy)->getBitWidth();
380 rv.IntVal = APInt(BitWidth, ((bool(*)())(intptr_t)FPtr)());
381 else if (BitWidth <= 8)
382 rv.IntVal = APInt(BitWidth, ((char(*)())(intptr_t)FPtr)());
383 else if (BitWidth <= 16)
384 rv.IntVal = APInt(BitWidth, ((short(*)())(intptr_t)FPtr)());
385 else if (BitWidth <= 32)
386 rv.IntVal = APInt(BitWidth, ((int(*)())(intptr_t)FPtr)());
387 else if (BitWidth <= 64)
388 rv.IntVal = APInt(BitWidth, ((int64_t(*)())(intptr_t)FPtr)());
390 assert(0 && "Integer types > 64 bits not supported");
394 rv.IntVal = APInt(32, ((int(*)())(intptr_t)FPtr)());
396 case Type::FloatTyID:
397 rv.FloatVal = ((float(*)())(intptr_t)FPtr)();
399 case Type::DoubleTyID:
400 rv.DoubleVal = ((double(*)())(intptr_t)FPtr)();
402 case Type::X86_FP80TyID:
403 case Type::FP128TyID:
404 case Type::PPC_FP128TyID:
405 assert(0 && "long double not supported yet");
407 case Type::PointerTyID:
408 return PTOGV(((void*(*)())(intptr_t)FPtr)());
412 // Okay, this is not one of our quick and easy cases. Because we don't have a
413 // full FFI, we have to codegen a nullary stub function that just calls the
414 // function we are interested in, passing in constants for all of the
415 // arguments. Make this function and return.
417 // First, create the function.
418 FunctionType *STy=FunctionType::get(RetTy, std::vector<const Type*>(), false);
419 Function *Stub = Function::Create(STy, Function::InternalLinkage, "",
422 // Insert a basic block.
423 BasicBlock *StubBB = BasicBlock::Create("", Stub);
425 // Convert all of the GenericValue arguments over to constants. Note that we
426 // currently don't support varargs.
427 SmallVector<Value*, 8> Args;
428 for (unsigned i = 0, e = ArgValues.size(); i != e; ++i) {
430 const Type *ArgTy = FTy->getParamType(i);
431 const GenericValue &AV = ArgValues[i];
432 switch (ArgTy->getTypeID()) {
433 default: assert(0 && "Unknown argument type for function call!");
434 case Type::IntegerTyID:
435 C = ConstantInt::get(AV.IntVal);
437 case Type::FloatTyID:
438 C = ConstantFP::get(APFloat(AV.FloatVal));
440 case Type::DoubleTyID:
441 C = ConstantFP::get(APFloat(AV.DoubleVal));
443 case Type::PPC_FP128TyID:
444 case Type::X86_FP80TyID:
445 case Type::FP128TyID:
446 C = ConstantFP::get(APFloat(AV.IntVal));
448 case Type::PointerTyID:
449 void *ArgPtr = GVTOP(AV);
450 if (sizeof(void*) == 4)
451 C = ConstantInt::get(Type::Int32Ty, (int)(intptr_t)ArgPtr);
453 C = ConstantInt::get(Type::Int64Ty, (intptr_t)ArgPtr);
454 C = ConstantExpr::getIntToPtr(C, ArgTy); // Cast the integer to pointer
460 CallInst *TheCall = CallInst::Create(F, Args.begin(), Args.end(),
462 TheCall->setCallingConv(F->getCallingConv());
463 TheCall->setTailCall();
464 if (TheCall->getType() != Type::VoidTy)
465 ReturnInst::Create(TheCall, StubBB); // Return result of the call.
467 ReturnInst::Create(StubBB); // Just return void.
469 // Finally, return the value returned by our nullary stub function.
470 return runFunction(Stub, std::vector<GenericValue>());
473 /// runJITOnFunction - Run the FunctionPassManager full of
474 /// just-in-time compilation passes on F, hopefully filling in
475 /// GlobalAddress[F] with the address of F's machine code.
477 void JIT::runJITOnFunction(Function *F) {
478 static bool isAlreadyCodeGenerating = false;
480 MutexGuard locked(lock);
481 assert(!isAlreadyCodeGenerating && "Error: Recursive compilation detected!");
484 isAlreadyCodeGenerating = true;
485 jitstate->getPM(locked).run(*F);
486 isAlreadyCodeGenerating = false;
488 // If the function referred to a global variable that had not yet been
489 // emitted, it allocates memory for the global, but doesn't emit it yet. Emit
490 // all of these globals now.
491 while (!jitstate->getPendingGlobals(locked).empty()) {
492 const GlobalVariable *GV = jitstate->getPendingGlobals(locked).back();
493 jitstate->getPendingGlobals(locked).pop_back();
494 EmitGlobalVariable(GV);
498 /// getPointerToFunction - This method is used to get the address of the
499 /// specified function, compiling it if neccesary.
501 void *JIT::getPointerToFunction(Function *F) {
503 if (void *Addr = getPointerToGlobalIfAvailable(F))
504 return Addr; // Check if function already code gen'd
506 // Make sure we read in the function if it exists in this Module.
507 if (F->hasNotBeenReadFromBitcode()) {
508 // Determine the module provider this function is provided by.
509 Module *M = F->getParent();
510 ModuleProvider *MP = 0;
511 for (unsigned i = 0, e = Modules.size(); i != e; ++i) {
512 if (Modules[i]->getModule() == M) {
517 assert(MP && "Function isn't in a module we know about!");
519 std::string ErrorMsg;
520 if (MP->materializeFunction(F, &ErrorMsg)) {
521 cerr << "Error reading function '" << F->getName()
522 << "' from bitcode file: " << ErrorMsg << "\n";
526 // Now retry to get the address.
527 if (void *Addr = getPointerToGlobalIfAvailable(F))
531 MutexGuard locked(lock);
533 if (F->isDeclaration()) {
534 bool AbortOnFailure = F->getLinkage() != GlobalValue::ExternalWeakLinkage;
535 void *Addr = getPointerToNamedFunction(F->getName(), AbortOnFailure);
536 addGlobalMapping(F, Addr);
542 void *Addr = getPointerToGlobalIfAvailable(F);
543 assert(Addr && "Code generation didn't add function to GlobalAddress table!");
547 /// getOrEmitGlobalVariable - Return the address of the specified global
548 /// variable, possibly emitting it to memory if needed. This is used by the
550 void *JIT::getOrEmitGlobalVariable(const GlobalVariable *GV) {
551 MutexGuard locked(lock);
553 void *Ptr = getPointerToGlobalIfAvailable(GV);
556 // If the global is external, just remember the address.
557 if (GV->isDeclaration()) {
558 #if HAVE___DSO_HANDLE
559 if (GV->getName() == "__dso_handle")
560 return (void*)&__dso_handle;
562 Ptr = sys::DynamicLibrary::SearchForAddressOfSymbol(GV->getName().c_str());
564 cerr << "Could not resolve external global address: "
565 << GV->getName() << "\n";
567 addGlobalMapping(GV, Ptr);
570 // GlobalVariable's which are not "constant" will cause trouble in a server
571 // situation. It's returned in the same block of memory as code which may
573 if (isGVCompilationDisabled() && !GV->isConstant()) {
574 cerr << "Compilation of non-internal GlobalValue is disabled!\n";
577 // If the global hasn't been emitted to memory yet, allocate space and
578 // emit it into memory. It goes in the same array as the generated
579 // code, jump tables, etc.
580 const Type *GlobalType = GV->getType()->getElementType();
581 size_t S = getTargetData()->getTypePaddedSize(GlobalType);
582 size_t A = getTargetData()->getPreferredAlignment(GV);
583 if (GV->isThreadLocal()) {
584 MutexGuard locked(lock);
585 Ptr = TJI.allocateThreadLocalMemory(S);
586 } else if (TJI.allocateSeparateGVMemory()) {
590 // Allocate S+A bytes of memory, then use an aligned pointer within that
593 unsigned MisAligned = ((intptr_t)Ptr & (A-1));
594 Ptr = (char*)Ptr + (MisAligned ? (A-MisAligned) : 0);
597 Ptr = MCE->allocateSpace(S, A);
599 addGlobalMapping(GV, Ptr);
600 EmitGlobalVariable(GV);
605 /// recompileAndRelinkFunction - This method is used to force a function
606 /// which has already been compiled, to be compiled again, possibly
607 /// after it has been modified. Then the entry to the old copy is overwritten
608 /// with a branch to the new copy. If there was no old copy, this acts
609 /// just like JIT::getPointerToFunction().
611 void *JIT::recompileAndRelinkFunction(Function *F) {
612 void *OldAddr = getPointerToGlobalIfAvailable(F);
614 // If it's not already compiled there is no reason to patch it up.
615 if (OldAddr == 0) { return getPointerToFunction(F); }
617 // Delete the old function mapping.
618 addGlobalMapping(F, 0);
620 // Recodegen the function
623 // Update state, forward the old function to the new function.
624 void *Addr = getPointerToGlobalIfAvailable(F);
625 assert(Addr && "Code generation didn't add function to GlobalAddress table!");
626 TJI.replaceMachineCodeForFunction(OldAddr, Addr);
630 /// getMemoryForGV - This method abstracts memory allocation of global
631 /// variable so that the JIT can allocate thread local variables depending
634 char* JIT::getMemoryForGV(const GlobalVariable* GV) {
635 const Type *ElTy = GV->getType()->getElementType();
636 size_t GVSize = (size_t)getTargetData()->getTypePaddedSize(ElTy);
637 if (GV->isThreadLocal()) {
638 MutexGuard locked(lock);
639 return TJI.allocateThreadLocalMemory(GVSize);
641 return new char[GVSize];