1 //===-- RuntimeDyld.h - Run-time dynamic linker for MC-JIT ------*- C++ -*-===//
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 // Implementation of the MC-JIT runtime dynamic linker.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "dyld"
15 #include "llvm/ADT/OwningPtr.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/StringMap.h"
18 #include "llvm/ADT/StringRef.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/Twine.h"
21 #include "llvm/ExecutionEngine/RuntimeDyld.h"
22 #include "llvm/Object/MachOObject.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/ErrorHandling.h"
25 #include "llvm/Support/Format.h"
26 #include "llvm/Support/Memory.h"
27 #include "llvm/Support/MemoryBuffer.h"
28 #include "llvm/Support/system_error.h"
29 #include "llvm/Support/raw_ostream.h"
31 using namespace llvm::object;
33 // Empty out-of-line virtual destructor as the key function.
34 RTDyldMemoryManager::~RTDyldMemoryManager() {}
37 class RuntimeDyldImpl {
41 // The MemoryManager to load objects into.
42 RTDyldMemoryManager *MemMgr;
45 // For each function, we have a MemoryBlock of it's instruction data.
46 StringMap<sys::MemoryBlock> Functions;
48 // Master symbol table. As modules are loaded and external symbols are
49 // resolved, their addresses are stored here.
50 StringMap<uint64_t> SymbolTable;
52 // FIXME: Should have multiple data blocks, one for each loaded chunk of
54 sys::MemoryBlock Data;
59 // Set the error state and record an error string.
60 bool Error(const Twine &Msg) {
66 void extractFunction(StringRef Name, uint8_t *StartAddress,
68 bool resolveRelocation(uint32_t BaseSection, macho::RelocationEntry RE,
69 SmallVectorImpl<void *> &SectionBases,
70 SmallVectorImpl<StringRef> &SymbolNames);
71 bool resolveX86_64Relocation(intptr_t Address, intptr_t Value, bool isPCRel,
72 unsigned Type, unsigned Size);
73 bool resolveARMRelocation(intptr_t Address, intptr_t Value, bool isPCRel,
74 unsigned Type, unsigned Size);
76 bool loadSegment32(const MachOObject *Obj,
77 const MachOObject::LoadCommandInfo *SegmentLCI,
78 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC);
79 bool loadSegment64(const MachOObject *Obj,
80 const MachOObject::LoadCommandInfo *SegmentLCI,
81 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC);
84 RuntimeDyldImpl(RTDyldMemoryManager *mm) : MemMgr(mm), HasError(false) {}
86 bool loadObject(MemoryBuffer *InputBuffer);
88 void *getSymbolAddress(StringRef Name) {
89 // FIXME: Just look up as a function for now. Overly simple of course.
91 return Functions.lookup(Name).base();
94 sys::MemoryBlock getMemoryBlock() { return Data; }
96 // Is the linker in an error state?
97 bool hasError() { return HasError; }
99 // Mark the error condition as handled and continue.
100 void clearError() { HasError = false; }
102 // Get the error message.
103 StringRef getErrorString() { return ErrorStr; }
106 void RuntimeDyldImpl::extractFunction(StringRef Name, uint8_t *StartAddress,
107 uint8_t *EndAddress) {
108 // Allocate memory for the function via the memory manager.
109 uintptr_t Size = EndAddress - StartAddress + 1;
110 uint8_t *Mem = MemMgr->startFunctionBody(Name.data(), Size);
111 assert(Size >= (uint64_t)(EndAddress - StartAddress + 1) &&
112 "Memory manager failed to allocate enough memory!");
113 // Copy the function payload into the memory block.
114 memcpy(Mem, StartAddress, EndAddress - StartAddress + 1);
115 MemMgr->endFunctionBody(Name.data(), Mem, Mem + Size);
116 // Remember where we put it.
117 Functions[Name] = sys::MemoryBlock(Mem, Size);
118 DEBUG(dbgs() << " allocated to " << Mem << "\n");
121 bool RuntimeDyldImpl::
122 resolveRelocation(uint32_t BaseSection, macho::RelocationEntry RE,
123 SmallVectorImpl<void *> &SectionBases,
124 SmallVectorImpl<StringRef> &SymbolNames) {
125 // struct relocation_info {
126 // int32_t r_address;
127 // uint32_t r_symbolnum:24,
133 uint32_t SymbolNum = RE.Word1 & 0xffffff; // 24-bit value
134 bool isPCRel = (RE.Word1 >> 24) & 1;
135 unsigned Log2Size = (RE.Word1 >> 25) & 3;
136 bool isExtern = (RE.Word1 >> 27) & 1;
137 unsigned Type = (RE.Word1 >> 28) & 0xf;
138 if (RE.Word0 & macho::RF_Scattered)
139 return Error("NOT YET IMPLEMENTED: scattered relocations.");
141 // The address requiring a relocation.
142 intptr_t Address = (intptr_t)SectionBases[BaseSection] + RE.Word0;
144 // Figure out the target address of the relocation. If isExtern is true,
145 // this relocation references the symbol table, otherwise it references
146 // a section in the same object, numbered from 1 through NumSections
147 // (SectionBases is [0, NumSections-1]).
150 StringRef Name = SymbolNames[SymbolNum];
151 if (SymbolTable.lookup(Name)) {
152 // The symbol is in our symbol table, so we can resolve it directly.
153 Value = (intptr_t)SymbolTable[Name];
155 return Error("NOT YET IMPLEMENTED: relocations to pre-compiled code.");
157 DEBUG(dbgs() << "Resolve relocation(" << Type << ") from '" << Name
158 << "' to " << format("0x%x", Address) << ".\n");
160 // For non-external relocations, the SymbolNum is actual a section number
161 // as described above.
162 Value = (intptr_t)SectionBases[SymbolNum - 1];
165 unsigned Size = 1 << Log2Size;
167 default: assert(0 && "Unsupported CPU type!");
168 case mach::CTM_x86_64:
169 return resolveX86_64Relocation(Address, Value, isPCRel, Type, Size);
171 return resolveARMRelocation(Address, Value, isPCRel, Type, Size);
173 llvm_unreachable("");
176 bool RuntimeDyldImpl::resolveX86_64Relocation(intptr_t Address, intptr_t Value,
177 bool isPCRel, unsigned Type,
179 // If the relocation is PC-relative, the value to be encoded is the
180 // pointer difference.
182 // FIXME: It seems this value needs to be adjusted by 4 for an effective PC
183 // address. Is that expected? Only for branches, perhaps?
184 Value -= Address + 4;
188 llvm_unreachable("Invalid relocation type!");
189 case macho::RIT_X86_64_Unsigned:
190 case macho::RIT_X86_64_Branch: {
191 // Mask in the target value a byte at a time (we don't have an alignment
192 // guarantee for the target address, so this is safest).
193 uint8_t *p = (uint8_t*)Address;
194 for (unsigned i = 0; i < Size; ++i) {
195 *p++ = (uint8_t)Value;
200 case macho::RIT_X86_64_Signed:
201 case macho::RIT_X86_64_GOTLoad:
202 case macho::RIT_X86_64_GOT:
203 case macho::RIT_X86_64_Subtractor:
204 case macho::RIT_X86_64_Signed1:
205 case macho::RIT_X86_64_Signed2:
206 case macho::RIT_X86_64_Signed4:
207 case macho::RIT_X86_64_TLV:
208 return Error("Relocation type not implemented yet!");
213 bool RuntimeDyldImpl::resolveARMRelocation(intptr_t Address, intptr_t Value,
214 bool isPCRel, unsigned Type,
216 // If the relocation is PC-relative, the value to be encoded is the
217 // pointer difference.
220 // ARM PCRel relocations have an effective-PC offset of two instructions
221 // (four bytes in Thumb mode, 8 bytes in ARM mode).
222 // FIXME: For now, assume ARM mode.
228 case macho::RIT_Vanilla: {
229 llvm_unreachable("Invalid relocation type!");
230 // Mask in the target value a byte at a time (we don't have an alignment
231 // guarantee for the target address, so this is safest).
232 uint8_t *p = (uint8_t*)Address;
233 for (unsigned i = 0; i < Size; ++i) {
234 *p++ = (uint8_t)Value;
239 case macho::RIT_Pair:
240 case macho::RIT_Difference:
241 case macho::RIT_ARM_LocalDifference:
242 case macho::RIT_ARM_PreboundLazyPointer:
243 case macho::RIT_ARM_Branch24Bit: {
244 // Mask the value into the target address. We know instructions are
245 // 32-bit aligned, so we can do it all at once.
246 uint32_t *p = (uint32_t*)Address;
247 // The low two bits of the value are not encoded.
249 // Mask the value to 24 bits.
251 // FIXME: If the destination is a Thumb function (and the instruction
252 // is a non-predicated BL instruction), we need to change it to a BLX
253 // instruction instead.
255 // Insert the value into the instruction.
256 *p = (*p & ~0xffffff) | Value;
259 case macho::RIT_ARM_ThumbBranch22Bit:
260 case macho::RIT_ARM_ThumbBranch32Bit:
261 case macho::RIT_ARM_Half:
262 case macho::RIT_ARM_HalfDifference:
263 return Error("Relocation type not implemented yet!");
268 bool RuntimeDyldImpl::
269 loadSegment32(const MachOObject *Obj,
270 const MachOObject::LoadCommandInfo *SegmentLCI,
271 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC) {
272 InMemoryStruct<macho::SegmentLoadCommand> Segment32LC;
273 Obj->ReadSegmentLoadCommand(*SegmentLCI, Segment32LC);
275 return Error("unable to load segment load command");
277 for (unsigned SectNum = 0; SectNum != Segment32LC->NumSections; ++SectNum) {
278 InMemoryStruct<macho::Section> Sect;
279 Obj->ReadSection(*SegmentLCI, SectNum, Sect);
281 return Error("unable to load section: '" + Twine(SectNum) + "'");
283 // FIXME: Improve check.
284 if (Sect->Flags != 0x80000400)
285 return Error("unsupported section type!");
287 // Address and names of symbols in the section.
288 typedef std::pair<uint64_t, StringRef> SymbolEntry;
289 SmallVector<SymbolEntry, 32> Symbols;
290 for (unsigned i = 0; i != SymtabLC->NumSymbolTableEntries; ++i) {
291 InMemoryStruct<macho::SymbolTableEntry> STE;
292 Obj->ReadSymbolTableEntry(SymtabLC->SymbolTableOffset, i, STE);
294 return Error("unable to read symbol: '" + Twine(i) + "'");
295 if (STE->SectionIndex > Segment32LC->NumSections)
296 return Error("invalid section index for symbol: '" + Twine() + "'");
298 // Just skip symbols not defined in this section.
299 if (STE->SectionIndex - 1 != SectNum)
302 // Get the symbol name.
303 StringRef Name = Obj->getStringAtIndex(STE->StringIndex);
305 // FIXME: Check the symbol type and flags.
306 if (STE->Type != 0xF) // external, defined in this section.
307 return Error("unexpected symbol type!");
308 if (STE->Flags != 0x0)
309 return Error("unexpected symbol type!");
311 uint64_t BaseAddress = Sect->Address;
312 uint64_t Address = BaseAddress + STE->Value;
314 // Remember the symbol.
315 Symbols.push_back(SymbolEntry(Address, Name));
317 DEBUG(dbgs() << "Function sym: '" << Name << "' @ " << Address << "\n");
319 // Sort the symbols by address, just in case they didn't come in that
321 array_pod_sort(Symbols.begin(), Symbols.end());
323 // Extract the function data.
324 uint8_t *Base = (uint8_t*)Obj->getData(Segment32LC->FileOffset,
325 Segment32LC->FileSize).data();
326 for (unsigned i = 0, e = Symbols.size() - 1; i != e; ++i) {
327 uint64_t StartOffset = Symbols[i].first;
328 uint64_t EndOffset = Symbols[i + 1].first - 1;
329 DEBUG(dbgs() << "Extracting function: " << Symbols[i].second
330 << " from [" << StartOffset << ", " << EndOffset << "]\n");
331 extractFunction(Symbols[i].second, Base + StartOffset, Base + EndOffset);
333 // The last symbol we do after since the end address is calculated
334 // differently because there is no next symbol to reference.
335 uint64_t StartOffset = Symbols[Symbols.size() - 1].first;
336 uint64_t EndOffset = Sect->Size - 1;
337 DEBUG(dbgs() << "Extracting function: " << Symbols[Symbols.size()-1].second
338 << " from [" << StartOffset << ", " << EndOffset << "]\n");
339 extractFunction(Symbols[Symbols.size()-1].second,
340 Base + StartOffset, Base + EndOffset);
347 bool RuntimeDyldImpl::
348 loadSegment64(const MachOObject *Obj,
349 const MachOObject::LoadCommandInfo *SegmentLCI,
350 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC) {
351 InMemoryStruct<macho::Segment64LoadCommand> Segment64LC;
352 Obj->ReadSegment64LoadCommand(*SegmentLCI, Segment64LC);
354 return Error("unable to load segment load command");
356 for (unsigned SectNum = 0; SectNum != Segment64LC->NumSections; ++SectNum) {
357 InMemoryStruct<macho::Section64> Sect;
358 Obj->ReadSection64(*SegmentLCI, SectNum, Sect);
360 return Error("unable to load section: '" + Twine(SectNum) + "'");
362 // FIXME: Improve check.
363 if (Sect->Flags != 0x80000400)
364 return Error("unsupported section type!");
366 // Address and names of symbols in the section.
367 typedef std::pair<uint64_t, StringRef> SymbolEntry;
368 SmallVector<SymbolEntry, 64> Symbols;
369 for (unsigned i = 0; i != SymtabLC->NumSymbolTableEntries; ++i) {
370 InMemoryStruct<macho::Symbol64TableEntry> STE;
371 Obj->ReadSymbol64TableEntry(SymtabLC->SymbolTableOffset, i, STE);
373 return Error("unable to read symbol: '" + Twine(i) + "'");
374 if (STE->SectionIndex > Segment64LC->NumSections)
375 return Error("invalid section index for symbol: '" + Twine() + "'");
377 // Just skip symbols not defined in this section.
378 if (STE->SectionIndex - 1 != SectNum)
381 // Get the symbol name.
382 StringRef Name = Obj->getStringAtIndex(STE->StringIndex);
384 // FIXME: Check the symbol type and flags.
385 if (STE->Type != 0xF) // external, defined in this section.
386 return Error("unexpected symbol type!");
387 if (STE->Flags != 0x0)
388 return Error("unexpected symbol type!");
390 uint64_t BaseAddress = Sect->Address;
391 uint64_t Address = BaseAddress + STE->Value;
393 // Remember the symbol.
394 Symbols.push_back(SymbolEntry(Address, Name));
396 DEBUG(dbgs() << "Function sym: '" << Name << "' @ " << Address << "\n");
398 // Sort the symbols by address, just in case they didn't come in that
400 array_pod_sort(Symbols.begin(), Symbols.end());
402 // Extract the function data.
403 uint8_t *Base = (uint8_t*)Obj->getData(Segment64LC->FileOffset,
404 Segment64LC->FileSize).data();
405 for (unsigned i = 0, e = Symbols.size() - 1; i != e; ++i) {
406 uint64_t StartOffset = Symbols[i].first;
407 uint64_t EndOffset = Symbols[i + 1].first - 1;
408 DEBUG(dbgs() << "Extracting function: " << Symbols[i].second
409 << " from [" << StartOffset << ", " << EndOffset << "]\n");
410 extractFunction(Symbols[i].second, Base + StartOffset, Base + EndOffset);
412 // The last symbol we do after since the end address is calculated
413 // differently because there is no next symbol to reference.
414 uint64_t StartOffset = Symbols[Symbols.size() - 1].first;
415 uint64_t EndOffset = Sect->Size - 1;
416 DEBUG(dbgs() << "Extracting function: " << Symbols[Symbols.size()-1].second
417 << " from [" << StartOffset << ", " << EndOffset << "]\n");
418 extractFunction(Symbols[Symbols.size()-1].second,
419 Base + StartOffset, Base + EndOffset);
425 bool RuntimeDyldImpl::loadObject(MemoryBuffer *InputBuffer) {
426 // If the linker is in an error state, don't do anything.
429 // Load the Mach-O wrapper object.
430 std::string ErrorStr;
431 OwningPtr<MachOObject> Obj(
432 MachOObject::LoadFromBuffer(InputBuffer, &ErrorStr));
434 return Error("unable to load object: '" + ErrorStr + "'");
436 // Get the CPU type information from the header.
437 const macho::Header &Header = Obj->getHeader();
439 // FIXME: Error checking that the loaded object is compatible with
440 // the system we're running on.
441 CPUType = Header.CPUType;
442 CPUSubtype = Header.CPUSubtype;
444 // Validate that the load commands match what we expect.
445 const MachOObject::LoadCommandInfo *SegmentLCI = 0, *SymtabLCI = 0,
447 for (unsigned i = 0; i != Header.NumLoadCommands; ++i) {
448 const MachOObject::LoadCommandInfo &LCI = Obj->getLoadCommandInfo(i);
449 switch (LCI.Command.Type) {
450 case macho::LCT_Segment:
451 case macho::LCT_Segment64:
453 return Error("unexpected input object (multiple segments)");
456 case macho::LCT_Symtab:
458 return Error("unexpected input object (multiple symbol tables)");
461 case macho::LCT_Dysymtab:
463 return Error("unexpected input object (multiple symbol tables)");
467 return Error("unexpected input object (unexpected load command");
472 return Error("no symbol table found in object");
474 return Error("no symbol table found in object");
476 // Read and register the symbol table data.
477 InMemoryStruct<macho::SymtabLoadCommand> SymtabLC;
478 Obj->ReadSymtabLoadCommand(*SymtabLCI, SymtabLC);
480 return Error("unable to load symbol table load command");
481 Obj->RegisterStringTable(*SymtabLC);
483 // Read the dynamic link-edit information, if present (not present in static
486 InMemoryStruct<macho::DysymtabLoadCommand> DysymtabLC;
487 Obj->ReadDysymtabLoadCommand(*DysymtabLCI, DysymtabLC);
489 return Error("unable to load dynamic link-exit load command");
491 // FIXME: We don't support anything interesting yet.
492 // if (DysymtabLC->LocalSymbolsIndex != 0)
493 // return Error("NOT YET IMPLEMENTED: local symbol entries");
494 // if (DysymtabLC->ExternalSymbolsIndex != 0)
495 // return Error("NOT YET IMPLEMENTED: non-external symbol entries");
496 // if (DysymtabLC->UndefinedSymbolsIndex != SymtabLC->NumSymbolTableEntries)
497 // return Error("NOT YET IMPLEMENTED: undefined symbol entries");
500 // Load the segment load command.
501 if (SegmentLCI->Command.Type == macho::LCT_Segment) {
502 if (loadSegment32(Obj.get(), SegmentLCI, SymtabLC))
505 if (loadSegment64(Obj.get(), SegmentLCI, SymtabLC))
513 //===----------------------------------------------------------------------===//
514 // RuntimeDyld class implementation
515 RuntimeDyld::RuntimeDyld(RTDyldMemoryManager *MM) {
516 Dyld = new RuntimeDyldImpl(MM);
519 RuntimeDyld::~RuntimeDyld() {
523 bool RuntimeDyld::loadObject(MemoryBuffer *InputBuffer) {
524 return Dyld->loadObject(InputBuffer);
527 void *RuntimeDyld::getSymbolAddress(StringRef Name) {
528 return Dyld->getSymbolAddress(Name);
531 sys::MemoryBlock RuntimeDyld::getMemoryBlock() {
532 return Dyld->getMemoryBlock();
535 StringRef RuntimeDyld::getErrorString() {
536 return Dyld->getErrorString();
539 } // end namespace llvm