1 //===-- RuntimeDyldMachO.cpp - 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/StringRef.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "RuntimeDyldImpl.h"
20 using namespace llvm::object;
24 bool RuntimeDyldMachO::
25 resolveRelocation(uint8_t *Address, uint64_t Value, bool isPCRel,
26 unsigned Type, unsigned Size, int64_t Addend) {
27 // This just dispatches to the proper target specific routine.
29 default: assert(0 && "Unsupported CPU type!");
30 case mach::CTM_x86_64:
31 return resolveX86_64Relocation((uintptr_t)Address, (uintptr_t)Value,
32 isPCRel, Type, Size, Addend);
34 return resolveARMRelocation((uintptr_t)Address, (uintptr_t)Value,
35 isPCRel, Type, Size, Addend);
40 bool RuntimeDyldMachO::
41 resolveX86_64Relocation(uintptr_t Address, uintptr_t Value, bool isPCRel,
42 unsigned Type, unsigned Size, int64_t Addend) {
43 // If the relocation is PC-relative, the value to be encoded is the
44 // pointer difference.
46 // FIXME: It seems this value needs to be adjusted by 4 for an effective PC
47 // address. Is that expected? Only for branches, perhaps?
52 llvm_unreachable("Invalid relocation type!");
53 case macho::RIT_X86_64_Signed1:
54 case macho::RIT_X86_64_Signed2:
55 case macho::RIT_X86_64_Signed4:
56 case macho::RIT_X86_64_Signed:
57 case macho::RIT_X86_64_Unsigned:
58 case macho::RIT_X86_64_Branch: {
60 // Mask in the target value a byte at a time (we don't have an alignment
61 // guarantee for the target address, so this is safest).
62 uint8_t *p = (uint8_t*)Address;
63 for (unsigned i = 0; i < Size; ++i) {
64 *p++ = (uint8_t)Value;
69 case macho::RIT_X86_64_GOTLoad:
70 case macho::RIT_X86_64_GOT:
71 case macho::RIT_X86_64_Subtractor:
72 case macho::RIT_X86_64_TLV:
73 return Error("Relocation type not implemented yet!");
78 bool RuntimeDyldMachO::
79 resolveARMRelocation(uintptr_t Address, uintptr_t Value, bool isPCRel,
80 unsigned Type, unsigned Size, int64_t Addend) {
81 // If the relocation is PC-relative, the value to be encoded is the
82 // pointer difference.
85 // ARM PCRel relocations have an effective-PC offset of two instructions
86 // (four bytes in Thumb mode, 8 bytes in ARM mode).
87 // FIXME: For now, assume ARM mode.
93 llvm_unreachable("Invalid relocation type!");
94 case macho::RIT_Vanilla: {
95 llvm_unreachable("Invalid relocation type!");
96 // Mask in the target value a byte at a time (we don't have an alignment
97 // guarantee for the target address, so this is safest).
98 uint8_t *p = (uint8_t*)Address;
99 for (unsigned i = 0; i < Size; ++i) {
100 *p++ = (uint8_t)Value;
105 case macho::RIT_ARM_Branch24Bit: {
106 // Mask the value into the target address. We know instructions are
107 // 32-bit aligned, so we can do it all at once.
108 uint32_t *p = (uint32_t*)Address;
109 // The low two bits of the value are not encoded.
111 // Mask the value to 24 bits.
113 // FIXME: If the destination is a Thumb function (and the instruction
114 // is a non-predicated BL instruction), we need to change it to a BLX
115 // instruction instead.
117 // Insert the value into the instruction.
118 *p = (*p & ~0xffffff) | Value;
121 case macho::RIT_ARM_ThumbBranch22Bit:
122 case macho::RIT_ARM_ThumbBranch32Bit:
123 case macho::RIT_ARM_Half:
124 case macho::RIT_ARM_HalfDifference:
125 case macho::RIT_Pair:
126 case macho::RIT_Difference:
127 case macho::RIT_ARM_LocalDifference:
128 case macho::RIT_ARM_PreboundLazyPointer:
129 return Error("Relocation type not implemented yet!");
134 bool RuntimeDyldMachO::
135 loadSegment32(const MachOObject *Obj,
136 const MachOObject::LoadCommandInfo *SegmentLCI,
137 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC) {
138 // FIXME: This should really be combined w/ loadSegment64. Templatized
139 // function on the 32/64 datatypes maybe?
140 InMemoryStruct<macho::SegmentLoadCommand> SegmentLC;
141 Obj->ReadSegmentLoadCommand(*SegmentLCI, SegmentLC);
143 return Error("unable to load segment load command");
146 SmallVector<unsigned, 16> SectionMap;
147 for (unsigned SectNum = 0; SectNum != SegmentLC->NumSections; ++SectNum) {
148 InMemoryStruct<macho::Section> Sect;
149 Obj->ReadSection(*SegmentLCI, SectNum, Sect);
151 return Error("unable to load section: '" + Twine(SectNum) + "'");
153 // Allocate memory via the MM for the section.
155 uint32_t SectionID = Sections.size();
156 if (Sect->Flags != 0x80000400)
157 Buffer = MemMgr->allocateCodeSection(Sect->Size, Sect->Align, SectionID);
159 Buffer = MemMgr->allocateDataSection(Sect->Size, Sect->Align, SectionID);
161 DEBUG(dbgs() << "Loading "
162 << ((Sect->Flags == 0x80000400) ? "text" : "data")
163 << " (ID #" << SectionID << ")"
164 << " '" << Sect->SegmentName << ","
165 << Sect->Name << "' of size " << Sect->Size
166 << " to address " << Buffer << ".\n");
168 // Copy the payload from the object file into the allocated buffer.
169 uint8_t *Base = (uint8_t*)Obj->getData(SegmentLC->FileOffset,
170 SegmentLC->FileSize).data();
171 memcpy(Buffer, Base + Sect->Address, Sect->Size);
173 // Remember what got allocated for this SectionID.
174 Sections.push_back(sys::MemoryBlock(Buffer, Sect->Size));
175 SectionLocalMemToID[Buffer] = SectionID;
177 // By default, the load address of a section is its memory buffer.
178 SectionLoadAddress.push_back((uint64_t)Buffer);
180 // Keep a map of object file section numbers to corresponding SectionIDs
181 // while processing the file.
182 SectionMap.push_back(SectionID);
185 // Process the symbol table.
186 SmallVector<StringRef, 64> SymbolNames;
187 processSymbols32(Obj, SectionMap, SymbolNames, SymtabLC);
189 // Process the relocations for each section we're loading.
190 Relocations.grow(Relocations.size() + SegmentLC->NumSections);
191 for (unsigned SectNum = 0; SectNum != SegmentLC->NumSections; ++SectNum) {
192 InMemoryStruct<macho::Section> Sect;
193 Obj->ReadSection(*SegmentLCI, SectNum, Sect);
195 return Error("unable to load section: '" + Twine(SectNum) + "'");
196 for (unsigned j = 0; j != Sect->NumRelocationTableEntries; ++j) {
197 InMemoryStruct<macho::RelocationEntry> RE;
198 Obj->ReadRelocationEntry(Sect->RelocationTableOffset, j, RE);
199 if (RE->Word0 & macho::RF_Scattered)
200 return Error("NOT YET IMPLEMENTED: scattered relocations.");
201 // Word0 of the relocation is the offset into the section where the
202 // relocation should be applied. We need to translate that into an
203 // offset into a function since that's our atom.
204 uint32_t Offset = RE->Word0;
205 bool isExtern = (RE->Word1 >> 27) & 1;
207 // FIXME: Get the relocation addend from the target address.
208 // FIXME: VERY imporant for internal relocations.
210 // Figure out the source symbol of the relocation. If isExtern is true,
211 // this relocation references the symbol table, otherwise it references
212 // a section in the same object, numbered from 1 through NumSections
213 // (SectionBases is [0, NumSections-1]).
214 uint32_t SourceNum = RE->Word1 & 0xffffff; // 24-bit value
216 assert(SourceNum > 0 && "Invalid relocation section number!");
217 unsigned SectionID = SectionMap[SourceNum - 1];
218 unsigned TargetID = SectionMap[SectNum];
219 DEBUG(dbgs() << "Internal relocation at Section #"
220 << TargetID << " + " << Offset
222 << SectionID << " (Word1: "
223 << format("0x%x", RE->Word1) << ")\n");
225 // Store the relocation information. It will get resolved when
226 // the section addresses are assigned.
227 Relocations[SectionID].push_back(RelocationEntry(TargetID,
232 StringRef SourceName = SymbolNames[SourceNum];
234 // Now store the relocation information. Associate it with the source
235 // symbol. Just add it to the unresolved list and let the general
236 // path post-load resolve it if we know where the symbol is.
237 UnresolvedRelocations[SourceName].push_back(RelocationEntry(SectNum,
241 DEBUG(dbgs() << "Relocation at Section #" << SectNum << " + " << Offset
242 << " from '" << SourceName << "(Word1: "
243 << format("0x%x", RE->Word1) << ")\n");
248 // Resolve the addresses of any symbols that were defined in this segment.
249 for (int i = 0, e = SymbolNames.size(); i != e; ++i)
250 resolveSymbol(SymbolNames[i]);
256 bool RuntimeDyldMachO::
257 loadSegment64(const MachOObject *Obj,
258 const MachOObject::LoadCommandInfo *SegmentLCI,
259 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC) {
260 InMemoryStruct<macho::Segment64LoadCommand> Segment64LC;
261 Obj->ReadSegment64LoadCommand(*SegmentLCI, Segment64LC);
263 return Error("unable to load segment load command");
266 SmallVector<unsigned, 16> SectionMap;
267 for (unsigned SectNum = 0; SectNum != Segment64LC->NumSections; ++SectNum) {
268 InMemoryStruct<macho::Section64> Sect;
269 Obj->ReadSection64(*SegmentLCI, SectNum, Sect);
271 return Error("unable to load section: '" + Twine(SectNum) + "'");
273 // Allocate memory via the MM for the section.
275 uint32_t SectionID = Sections.size();
276 if (Sect->Flags != 0x80000400)
277 Buffer = MemMgr->allocateCodeSection(Sect->Size, Sect->Align, SectionID);
279 Buffer = MemMgr->allocateDataSection(Sect->Size, Sect->Align, SectionID);
281 DEBUG(dbgs() << "Loading "
282 << ((Sect->Flags == 0x80000400) ? "text" : "data")
283 << " (ID #" << SectionID << ")"
284 << " '" << Sect->SegmentName << ","
285 << Sect->Name << "' of size " << Sect->Size
286 << " to address " << Buffer << ".\n");
288 // Copy the payload from the object file into the allocated buffer.
289 uint8_t *Base = (uint8_t*)Obj->getData(Segment64LC->FileOffset,
290 Segment64LC->FileSize).data();
291 memcpy(Buffer, Base + Sect->Address, Sect->Size);
293 // Remember what got allocated for this SectionID.
294 Sections.push_back(sys::MemoryBlock(Buffer, Sect->Size));
295 SectionLocalMemToID[Buffer] = SectionID;
297 // By default, the load address of a section is its memory buffer.
298 SectionLoadAddress.push_back((uint64_t)Buffer);
300 // Keep a map of object file section numbers to corresponding SectionIDs
301 // while processing the file.
302 SectionMap.push_back(SectionID);
305 // Process the symbol table.
306 SmallVector<StringRef, 64> SymbolNames;
307 processSymbols64(Obj, SectionMap, SymbolNames, SymtabLC);
309 // Process the relocations for each section we're loading.
310 Relocations.grow(Relocations.size() + Segment64LC->NumSections);
311 for (unsigned SectNum = 0; SectNum != Segment64LC->NumSections; ++SectNum) {
312 InMemoryStruct<macho::Section64> Sect;
313 Obj->ReadSection64(*SegmentLCI, SectNum, Sect);
315 return Error("unable to load section: '" + Twine(SectNum) + "'");
316 for (unsigned j = 0; j != Sect->NumRelocationTableEntries; ++j) {
317 InMemoryStruct<macho::RelocationEntry> RE;
318 Obj->ReadRelocationEntry(Sect->RelocationTableOffset, j, RE);
319 if (RE->Word0 & macho::RF_Scattered)
320 return Error("NOT YET IMPLEMENTED: scattered relocations.");
321 // Word0 of the relocation is the offset into the section where the
322 // relocation should be applied. We need to translate that into an
323 // offset into a function since that's our atom.
324 uint32_t Offset = RE->Word0;
325 bool isExtern = (RE->Word1 >> 27) & 1;
327 // FIXME: Get the relocation addend from the target address.
328 // FIXME: VERY imporant for internal relocations.
330 // Figure out the source symbol of the relocation. If isExtern is true,
331 // this relocation references the symbol table, otherwise it references
332 // a section in the same object, numbered from 1 through NumSections
333 // (SectionBases is [0, NumSections-1]).
334 uint32_t SourceNum = RE->Word1 & 0xffffff; // 24-bit value
336 assert(SourceNum > 0 && "Invalid relocation section number!");
337 unsigned SectionID = SectionMap[SourceNum - 1];
338 unsigned TargetID = SectionMap[SectNum];
339 DEBUG(dbgs() << "Internal relocation at Section #"
340 << TargetID << " + " << Offset
342 << SectionID << " (Word1: "
343 << format("0x%x", RE->Word1) << ")\n");
345 // Store the relocation information. It will get resolved when
346 // the section addresses are assigned.
347 Relocations[SectionID].push_back(RelocationEntry(TargetID,
352 StringRef SourceName = SymbolNames[SourceNum];
354 // Now store the relocation information. Associate it with the source
355 // symbol. Just add it to the unresolved list and let the general
356 // path post-load resolve it if we know where the symbol is.
357 UnresolvedRelocations[SourceName].push_back(RelocationEntry(SectNum,
361 DEBUG(dbgs() << "Relocation at Section #" << SectNum << " + " << Offset
362 << " from '" << SourceName << "(Word1: "
363 << format("0x%x", RE->Word1) << ")\n");
368 // Resolve the addresses of any symbols that were defined in this segment.
369 for (int i = 0, e = SymbolNames.size(); i != e; ++i)
370 resolveSymbol(SymbolNames[i]);
375 bool RuntimeDyldMachO::
376 processSymbols32(const MachOObject *Obj,
377 SmallVectorImpl<unsigned> &SectionMap,
378 SmallVectorImpl<StringRef> &SymbolNames,
379 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC) {
380 // FIXME: Combine w/ processSymbols64. Factor 64/32 datatype and such.
381 for (unsigned i = 0; i != SymtabLC->NumSymbolTableEntries; ++i) {
382 InMemoryStruct<macho::SymbolTableEntry> STE;
383 Obj->ReadSymbolTableEntry(SymtabLC->SymbolTableOffset, i, STE);
385 return Error("unable to read symbol: '" + Twine(i) + "'");
386 // Get the symbol name.
387 StringRef Name = Obj->getStringAtIndex(STE->StringIndex);
388 SymbolNames.push_back(Name);
390 // FIXME: Check the symbol type and flags.
391 if (STE->Type != 0xF) // external, defined in this segment.
393 // Flags in the upper nibble we don't care about.
394 if ((STE->Flags & 0xf) != 0x0)
397 // Remember the symbol.
398 uint32_t SectionID = SectionMap[STE->SectionIndex - 1];
399 SymbolTable[Name] = SymbolLoc(SectionID, STE->Value);
401 DEBUG(dbgs() << "Symbol: '" << Name << "' @ "
402 << (getSectionAddress(SectionID) + STE->Value)
408 bool RuntimeDyldMachO::
409 processSymbols64(const MachOObject *Obj,
410 SmallVectorImpl<unsigned> &SectionMap,
411 SmallVectorImpl<StringRef> &SymbolNames,
412 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC) {
413 for (unsigned i = 0; i != SymtabLC->NumSymbolTableEntries; ++i) {
414 InMemoryStruct<macho::Symbol64TableEntry> STE;
415 Obj->ReadSymbol64TableEntry(SymtabLC->SymbolTableOffset, i, STE);
417 return Error("unable to read symbol: '" + Twine(i) + "'");
418 // Get the symbol name.
419 StringRef Name = Obj->getStringAtIndex(STE->StringIndex);
420 SymbolNames.push_back(Name);
422 // FIXME: Check the symbol type and flags.
423 if (STE->Type != 0xF) // external, defined in this segment.
425 // Flags in the upper nibble we don't care about.
426 if ((STE->Flags & 0xf) != 0x0)
429 // Remember the symbol.
430 uint32_t SectionID = SectionMap[STE->SectionIndex - 1];
431 SymbolTable[Name] = SymbolLoc(SectionID, STE->Value);
433 DEBUG(dbgs() << "Symbol: '" << Name << "' @ "
434 << (getSectionAddress(SectionID) + STE->Value)
440 // resolveSymbol - Resolve any relocations to the specified symbol if
441 // we know where it lives.
442 void RuntimeDyldMachO::resolveSymbol(StringRef Name) {
443 StringMap<SymbolLoc>::const_iterator Loc = SymbolTable.find(Name);
444 if (Loc == SymbolTable.end())
447 RelocationList &Relocs = UnresolvedRelocations[Name];
448 DEBUG(dbgs() << "Resolving symbol '" << Name << "'\n");
449 for (int i = 0, e = Relocs.size(); i != e; ++i) {
450 // Change the relocation to be section relative rather than symbol
451 // relative and move it to the resolved relocation list.
452 RelocationEntry Entry = Relocs[i];
453 Entry.Addend += Loc->second.second;
454 Relocations[Loc->second.first].push_back(Entry);
456 // FIXME: Keep a worklist of the relocations we've added so that we can
457 // resolve more selectively later.
461 bool RuntimeDyldMachO::loadObject(MemoryBuffer *InputBuffer) {
462 // If the linker is in an error state, don't do anything.
465 // Load the Mach-O wrapper object.
466 std::string ErrorStr;
467 OwningPtr<MachOObject> Obj(
468 MachOObject::LoadFromBuffer(InputBuffer, &ErrorStr));
470 return Error("unable to load object: '" + ErrorStr + "'");
472 // Get the CPU type information from the header.
473 const macho::Header &Header = Obj->getHeader();
475 // FIXME: Error checking that the loaded object is compatible with
476 // the system we're running on.
477 CPUType = Header.CPUType;
478 CPUSubtype = Header.CPUSubtype;
480 // Validate that the load commands match what we expect.
481 const MachOObject::LoadCommandInfo *SegmentLCI = 0, *SymtabLCI = 0,
483 for (unsigned i = 0; i != Header.NumLoadCommands; ++i) {
484 const MachOObject::LoadCommandInfo &LCI = Obj->getLoadCommandInfo(i);
485 switch (LCI.Command.Type) {
486 case macho::LCT_Segment:
487 case macho::LCT_Segment64:
489 return Error("unexpected input object (multiple segments)");
492 case macho::LCT_Symtab:
494 return Error("unexpected input object (multiple symbol tables)");
497 case macho::LCT_Dysymtab:
499 return Error("unexpected input object (multiple symbol tables)");
503 return Error("unexpected input object (unexpected load command");
508 return Error("no symbol table found in object");
510 return Error("no segments found in object");
512 // Read and register the symbol table data.
513 InMemoryStruct<macho::SymtabLoadCommand> SymtabLC;
514 Obj->ReadSymtabLoadCommand(*SymtabLCI, SymtabLC);
516 return Error("unable to load symbol table load command");
517 Obj->RegisterStringTable(*SymtabLC);
519 // Read the dynamic link-edit information, if present (not present in static
522 InMemoryStruct<macho::DysymtabLoadCommand> DysymtabLC;
523 Obj->ReadDysymtabLoadCommand(*DysymtabLCI, DysymtabLC);
525 return Error("unable to load dynamic link-exit load command");
527 // FIXME: We don't support anything interesting yet.
528 // if (DysymtabLC->LocalSymbolsIndex != 0)
529 // return Error("NOT YET IMPLEMENTED: local symbol entries");
530 // if (DysymtabLC->ExternalSymbolsIndex != 0)
531 // return Error("NOT YET IMPLEMENTED: non-external symbol entries");
532 // if (DysymtabLC->UndefinedSymbolsIndex != SymtabLC->NumSymbolTableEntries)
533 // return Error("NOT YET IMPLEMENTED: undefined symbol entries");
536 // Load the segment load command.
537 if (SegmentLCI->Command.Type == macho::LCT_Segment) {
538 if (loadSegment32(Obj.get(), SegmentLCI, SymtabLC))
541 if (loadSegment64(Obj.get(), SegmentLCI, SymtabLC))
545 // Assign the addresses of the sections from the object so that any
546 // relocations to them get set properly.
547 // FIXME: This is done directly from the client at the moment. We should
548 // default the values to the local storage, at least when the target arch
549 // is the same as the host arch.
554 // Assign an address to a symbol name and resolve all the relocations
555 // associated with it.
556 void RuntimeDyldMachO::reassignSectionAddress(unsigned SectionID,
558 // The address to use for relocation resolution is not
559 // the address of the local section buffer. We must be doing
560 // a remote execution environment of some sort. Re-apply any
561 // relocations referencing this section with the given address.
563 // Addr is a uint64_t because we can't assume the pointer width
564 // of the target is the same as that of the host. Just use a generic
565 // "big enough" type.
567 SectionLoadAddress[SectionID] = Addr;
569 RelocationList &Relocs = Relocations[SectionID];
570 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
571 RelocationEntry &RE = Relocs[i];
572 uint8_t *Target = (uint8_t*)Sections[RE.SectionID].base() + RE.Offset;
573 bool isPCRel = (RE.Data >> 24) & 1;
574 unsigned Type = (RE.Data >> 28) & 0xf;
575 unsigned Size = 1 << ((RE.Data >> 25) & 3);
577 DEBUG(dbgs() << "Resolving relocation at Section #" << RE.SectionID
578 << " + " << RE.Offset << " (" << format("%p", Target) << ")"
579 << " from Section #" << SectionID << " (" << format("%p", Addr) << ")"
580 << "(" << (isPCRel ? "pcrel" : "absolute")
581 << ", type: " << Type << ", Size: " << Size << ", Addend: "
582 << RE.Addend << ").\n");
584 resolveRelocation(Target, Addr, isPCRel, Type, Size, RE.Addend);
588 bool RuntimeDyldMachO::isKnownFormat(const MemoryBuffer *InputBuffer) {
589 StringRef Magic = InputBuffer->getBuffer().slice(0, 4);
590 if (Magic == "\xFE\xED\xFA\xCE") return true;
591 if (Magic == "\xCE\xFA\xED\xFE") return true;
592 if (Magic == "\xFE\xED\xFA\xCF") return true;
593 if (Magic == "\xCF\xFA\xED\xFE") return true;
597 } // end namespace llvm