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 "RuntimeDyldMachO.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: llvm_unreachable("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);
39 bool RuntimeDyldMachO::
40 resolveX86_64Relocation(uintptr_t Address, uintptr_t Value, bool isPCRel,
41 unsigned Type, unsigned Size, int64_t Addend) {
42 // If the relocation is PC-relative, the value to be encoded is the
43 // pointer difference.
45 // FIXME: It seems this value needs to be adjusted by 4 for an effective PC
46 // address. Is that expected? Only for branches, perhaps?
51 llvm_unreachable("Invalid relocation type!");
52 case macho::RIT_X86_64_Signed1:
53 case macho::RIT_X86_64_Signed2:
54 case macho::RIT_X86_64_Signed4:
55 case macho::RIT_X86_64_Signed:
56 case macho::RIT_X86_64_Unsigned:
57 case macho::RIT_X86_64_Branch: {
59 // Mask in the target value a byte at a time (we don't have an alignment
60 // guarantee for the target address, so this is safest).
61 uint8_t *p = (uint8_t*)Address;
62 for (unsigned i = 0; i < Size; ++i) {
63 *p++ = (uint8_t)Value;
68 case macho::RIT_X86_64_GOTLoad:
69 case macho::RIT_X86_64_GOT:
70 case macho::RIT_X86_64_Subtractor:
71 case macho::RIT_X86_64_TLV:
72 return Error("Relocation type not implemented yet!");
76 bool RuntimeDyldMachO::
77 resolveARMRelocation(uintptr_t Address, uintptr_t Value, bool isPCRel,
78 unsigned Type, unsigned Size, int64_t Addend) {
79 // If the relocation is PC-relative, the value to be encoded is the
80 // pointer difference.
83 // ARM PCRel relocations have an effective-PC offset of two instructions
84 // (four bytes in Thumb mode, 8 bytes in ARM mode).
85 // FIXME: For now, assume ARM mode.
91 llvm_unreachable("Invalid relocation type!");
92 case macho::RIT_Vanilla: {
93 // Mask in the target value a byte at a time (we don't have an alignment
94 // guarantee for the target address, so this is safest).
95 uint8_t *p = (uint8_t*)Address;
96 for (unsigned i = 0; i < Size; ++i) {
97 *p++ = (uint8_t)Value;
102 case macho::RIT_ARM_Branch24Bit: {
103 // Mask the value into the target address. We know instructions are
104 // 32-bit aligned, so we can do it all at once.
105 uint32_t *p = (uint32_t*)Address;
106 // The low two bits of the value are not encoded.
108 // Mask the value to 24 bits.
110 // FIXME: If the destination is a Thumb function (and the instruction
111 // is a non-predicated BL instruction), we need to change it to a BLX
112 // instruction instead.
114 // Insert the value into the instruction.
115 *p = (*p & ~0xffffff) | Value;
118 case macho::RIT_ARM_ThumbBranch22Bit:
119 case macho::RIT_ARM_ThumbBranch32Bit:
120 case macho::RIT_ARM_Half:
121 case macho::RIT_ARM_HalfDifference:
122 case macho::RIT_Pair:
123 case macho::RIT_Difference:
124 case macho::RIT_ARM_LocalDifference:
125 case macho::RIT_ARM_PreboundLazyPointer:
126 return Error("Relocation type not implemented yet!");
131 bool RuntimeDyldMachO::
132 loadSegment32(const MachOObject *Obj,
133 const MachOObject::LoadCommandInfo *SegmentLCI,
134 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC) {
135 // FIXME: This should really be combined w/ loadSegment64. Templatized
136 // function on the 32/64 datatypes maybe?
137 InMemoryStruct<macho::SegmentLoadCommand> SegmentLC;
138 Obj->ReadSegmentLoadCommand(*SegmentLCI, SegmentLC);
140 return Error("unable to load segment load command");
143 SmallVector<unsigned, 16> SectionMap;
144 for (unsigned SectNum = 0; SectNum != SegmentLC->NumSections; ++SectNum) {
145 InMemoryStruct<macho::Section> Sect;
146 Obj->ReadSection(*SegmentLCI, SectNum, Sect);
148 return Error("unable to load section: '" + Twine(SectNum) + "'");
150 // Allocate memory via the MM for the section.
152 uint32_t SectionID = Sections.size();
153 if (Sect->Flags == 0x80000400)
154 Buffer = MemMgr->allocateCodeSection(Sect->Size, Sect->Align, SectionID);
156 Buffer = MemMgr->allocateDataSection(Sect->Size, Sect->Align, SectionID);
158 DEBUG(dbgs() << "Loading "
159 << ((Sect->Flags == 0x80000400) ? "text" : "data")
160 << " (ID #" << SectionID << ")"
161 << " '" << Sect->SegmentName << ","
162 << Sect->Name << "' of size " << Sect->Size
163 << " to address " << Buffer << ".\n");
165 // Copy the payload from the object file into the allocated buffer.
166 uint8_t *Base = (uint8_t*)Obj->getData(SegmentLC->FileOffset,
167 SegmentLC->FileSize).data();
168 memcpy(Buffer, Base + Sect->Address, Sect->Size);
170 // Remember what got allocated for this SectionID.
171 Sections.push_back(sys::MemoryBlock(Buffer, Sect->Size));
172 SectionLocalMemToID[Buffer] = SectionID;
174 // By default, the load address of a section is its memory buffer.
175 SectionLoadAddress.push_back((uint64_t)Buffer);
177 // Keep a map of object file section numbers to corresponding SectionIDs
178 // while processing the file.
179 SectionMap.push_back(SectionID);
182 // Process the symbol table.
183 SmallVector<StringRef, 64> SymbolNames;
184 processSymbols32(Obj, SectionMap, SymbolNames, SymtabLC);
186 // Process the relocations for each section we're loading.
187 Relocations.grow(Relocations.size() + SegmentLC->NumSections);
188 for (unsigned SectNum = 0; SectNum != SegmentLC->NumSections; ++SectNum) {
189 InMemoryStruct<macho::Section> Sect;
190 Obj->ReadSection(*SegmentLCI, SectNum, Sect);
192 return Error("unable to load section: '" + Twine(SectNum) + "'");
193 for (unsigned j = 0; j != Sect->NumRelocationTableEntries; ++j) {
194 InMemoryStruct<macho::RelocationEntry> RE;
195 Obj->ReadRelocationEntry(Sect->RelocationTableOffset, j, RE);
196 if (RE->Word0 & macho::RF_Scattered)
197 return Error("NOT YET IMPLEMENTED: scattered relocations.");
198 // Word0 of the relocation is the offset into the section where the
199 // relocation should be applied. We need to translate that into an
200 // offset into a function since that's our atom.
201 uint32_t Offset = RE->Word0;
202 bool isExtern = (RE->Word1 >> 27) & 1;
204 // FIXME: Get the relocation addend from the target address.
205 // FIXME: VERY imporant for internal relocations.
207 // Figure out the source symbol of the relocation. If isExtern is true,
208 // this relocation references the symbol table, otherwise it references
209 // a section in the same object, numbered from 1 through NumSections
210 // (SectionBases is [0, NumSections-1]).
211 uint32_t SourceNum = RE->Word1 & 0xffffff; // 24-bit value
213 assert(SourceNum > 0 && "Invalid relocation section number!");
214 unsigned SectionID = SectionMap[SourceNum - 1];
215 unsigned TargetID = SectionMap[SectNum];
216 DEBUG(dbgs() << "Internal relocation at Section #"
217 << TargetID << " + " << Offset
219 << SectionID << " (Word1: "
220 << format("0x%x", RE->Word1) << ")\n");
222 // Store the relocation information. It will get resolved when
223 // the section addresses are assigned.
224 Relocations[SectionID].push_back(RelocationEntry(TargetID,
229 StringRef SourceName = SymbolNames[SourceNum];
231 // Now store the relocation information. Associate it with the source
232 // symbol. Just add it to the unresolved list and let the general
233 // path post-load resolve it if we know where the symbol is.
234 UnresolvedRelocations[SourceName].push_back(RelocationEntry(SectNum,
238 DEBUG(dbgs() << "Relocation at Section #" << SectNum << " + " << Offset
239 << " from '" << SourceName << "(Word1: "
240 << format("0x%x", RE->Word1) << ")\n");
245 // Resolve the addresses of any symbols that were defined in this segment.
246 for (int i = 0, e = SymbolNames.size(); i != e; ++i)
247 resolveSymbol(SymbolNames[i]);
253 bool RuntimeDyldMachO::
254 loadSegment64(const MachOObject *Obj,
255 const MachOObject::LoadCommandInfo *SegmentLCI,
256 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC) {
257 InMemoryStruct<macho::Segment64LoadCommand> Segment64LC;
258 Obj->ReadSegment64LoadCommand(*SegmentLCI, Segment64LC);
260 return Error("unable to load segment load command");
263 SmallVector<unsigned, 16> SectionMap;
264 for (unsigned SectNum = 0; SectNum != Segment64LC->NumSections; ++SectNum) {
265 InMemoryStruct<macho::Section64> Sect;
266 Obj->ReadSection64(*SegmentLCI, SectNum, Sect);
268 return Error("unable to load section: '" + Twine(SectNum) + "'");
270 // Allocate memory via the MM for the section.
272 uint32_t SectionID = Sections.size();
273 unsigned Align = 1 << Sect->Align; // .o file has log2 alignment.
274 if (Sect->Flags == 0x80000400)
275 Buffer = MemMgr->allocateCodeSection(Sect->Size, Align, SectionID);
277 Buffer = MemMgr->allocateDataSection(Sect->Size, Align, SectionID);
279 DEBUG(dbgs() << "Loading "
280 << ((Sect->Flags == 0x80000400) ? "text" : "data")
281 << " (ID #" << SectionID << ")"
282 << " '" << Sect->SegmentName << ","
283 << Sect->Name << "' of size " << Sect->Size
284 << " (align " << Align << ")"
285 << " to address " << Buffer << ".\n");
287 // Copy the payload from the object file into the allocated buffer.
288 uint8_t *Base = (uint8_t*)Obj->getData(Segment64LC->FileOffset,
289 Segment64LC->FileSize).data();
290 memcpy(Buffer, Base + Sect->Address, Sect->Size);
292 // Remember what got allocated for this SectionID.
293 Sections.push_back(sys::MemoryBlock(Buffer, Sect->Size));
294 SectionLocalMemToID[Buffer] = SectionID;
296 // By default, the load address of a section is its memory buffer.
297 SectionLoadAddress.push_back((uint64_t)Buffer);
299 // Keep a map of object file section numbers to corresponding SectionIDs
300 // while processing the file.
301 SectionMap.push_back(SectionID);
304 // Process the symbol table.
305 SmallVector<StringRef, 64> SymbolNames;
306 processSymbols64(Obj, SectionMap, SymbolNames, SymtabLC);
308 // Process the relocations for each section we're loading.
309 Relocations.grow(Relocations.size() + Segment64LC->NumSections);
310 for (unsigned SectNum = 0; SectNum != Segment64LC->NumSections; ++SectNum) {
311 InMemoryStruct<macho::Section64> Sect;
312 Obj->ReadSection64(*SegmentLCI, SectNum, Sect);
314 return Error("unable to load section: '" + Twine(SectNum) + "'");
315 for (unsigned j = 0; j != Sect->NumRelocationTableEntries; ++j) {
316 InMemoryStruct<macho::RelocationEntry> RE;
317 Obj->ReadRelocationEntry(Sect->RelocationTableOffset, j, RE);
318 if (RE->Word0 & macho::RF_Scattered)
319 return Error("NOT YET IMPLEMENTED: scattered relocations.");
320 // Word0 of the relocation is the offset into the section where the
321 // relocation should be applied. We need to translate that into an
322 // offset into a function since that's our atom.
323 uint32_t Offset = RE->Word0;
324 bool isExtern = (RE->Word1 >> 27) & 1;
326 // FIXME: Get the relocation addend from the target address.
327 // FIXME: VERY imporant for internal relocations.
329 // Figure out the source symbol of the relocation. If isExtern is true,
330 // this relocation references the symbol table, otherwise it references
331 // a section in the same object, numbered from 1 through NumSections
332 // (SectionBases is [0, NumSections-1]).
333 uint32_t SourceNum = RE->Word1 & 0xffffff; // 24-bit value
335 assert(SourceNum > 0 && "Invalid relocation section number!");
336 unsigned SectionID = SectionMap[SourceNum - 1];
337 unsigned TargetID = SectionMap[SectNum];
338 DEBUG(dbgs() << "Internal relocation at Section #"
339 << TargetID << " + " << Offset
341 << SectionID << " (Word1: "
342 << format("0x%x", RE->Word1) << ")\n");
344 // Store the relocation information. It will get resolved when
345 // the section addresses are assigned.
346 Relocations[SectionID].push_back(RelocationEntry(TargetID,
351 StringRef SourceName = SymbolNames[SourceNum];
353 // Now store the relocation information. Associate it with the source
354 // symbol. Just add it to the unresolved list and let the general
355 // path post-load resolve it if we know where the symbol is.
356 UnresolvedRelocations[SourceName].push_back(RelocationEntry(SectNum,
360 DEBUG(dbgs() << "Relocation at Section #" << SectNum << " + " << Offset
361 << " from '" << SourceName << "(Word1: "
362 << format("0x%x", RE->Word1) << ")\n");
367 // Resolve the addresses of any symbols that were defined in this segment.
368 for (int i = 0, e = SymbolNames.size(); i != e; ++i)
369 resolveSymbol(SymbolNames[i]);
374 bool RuntimeDyldMachO::
375 processSymbols32(const MachOObject *Obj,
376 SmallVectorImpl<unsigned> &SectionMap,
377 SmallVectorImpl<StringRef> &SymbolNames,
378 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC) {
379 // FIXME: Combine w/ processSymbols64. Factor 64/32 datatype and such.
380 for (unsigned i = 0; i != SymtabLC->NumSymbolTableEntries; ++i) {
381 InMemoryStruct<macho::SymbolTableEntry> STE;
382 Obj->ReadSymbolTableEntry(SymtabLC->SymbolTableOffset, i, STE);
384 return Error("unable to read symbol: '" + Twine(i) + "'");
385 // Get the symbol name.
386 StringRef Name = Obj->getStringAtIndex(STE->StringIndex);
387 SymbolNames.push_back(Name);
389 // FIXME: Check the symbol type and flags.
390 if (STE->Type != 0xF) // external, defined in this segment.
392 // Flags in the upper nibble we don't care about.
393 if ((STE->Flags & 0xf) != 0x0)
396 // Remember the symbol.
397 uint32_t SectionID = SectionMap[STE->SectionIndex - 1];
398 SymbolTable[Name] = SymbolLoc(SectionID, STE->Value);
400 DEBUG(dbgs() << "Symbol: '" << Name << "' @ "
401 << (getSectionAddress(SectionID) + STE->Value)
407 bool RuntimeDyldMachO::
408 processSymbols64(const MachOObject *Obj,
409 SmallVectorImpl<unsigned> &SectionMap,
410 SmallVectorImpl<StringRef> &SymbolNames,
411 const InMemoryStruct<macho::SymtabLoadCommand> &SymtabLC) {
412 for (unsigned i = 0; i != SymtabLC->NumSymbolTableEntries; ++i) {
413 InMemoryStruct<macho::Symbol64TableEntry> STE;
414 Obj->ReadSymbol64TableEntry(SymtabLC->SymbolTableOffset, i, STE);
416 return Error("unable to read symbol: '" + Twine(i) + "'");
417 // Get the symbol name.
418 StringRef Name = Obj->getStringAtIndex(STE->StringIndex);
419 SymbolNames.push_back(Name);
421 // FIXME: Check the symbol type and flags.
422 if (STE->Type != 0xF) // external, defined in this segment.
424 // Flags in the upper nibble we don't care about.
425 if ((STE->Flags & 0xf) != 0x0)
428 // Remember the symbol.
429 uint32_t SectionID = SectionMap[STE->SectionIndex - 1];
430 SymbolTable[Name] = SymbolLoc(SectionID, STE->Value);
432 DEBUG(dbgs() << "Symbol: '" << Name << "' @ "
433 << (getSectionAddress(SectionID) + STE->Value)
439 // resolveSymbol - Resolve any relocations to the specified symbol if
440 // we know where it lives.
441 void RuntimeDyldMachO::resolveSymbol(StringRef Name) {
442 StringMap<SymbolLoc>::const_iterator Loc = SymbolTable.find(Name);
443 if (Loc == SymbolTable.end())
446 RelocationList &Relocs = UnresolvedRelocations[Name];
447 DEBUG(dbgs() << "Resolving symbol '" << Name << "'\n");
448 for (int i = 0, e = Relocs.size(); i != e; ++i) {
449 // Change the relocation to be section relative rather than symbol
450 // relative and move it to the resolved relocation list.
451 RelocationEntry Entry = Relocs[i];
452 Entry.Addend += Loc->second.second;
453 Relocations[Loc->second.first].push_back(Entry);
455 // FIXME: Keep a worklist of the relocations we've added so that we can
456 // resolve more selectively later.
460 bool RuntimeDyldMachO::loadObject(MemoryBuffer *InputBuffer) {
461 // If the linker is in an error state, don't do anything.
464 // Load the Mach-O wrapper object.
465 std::string ErrorStr;
466 OwningPtr<MachOObject> Obj(
467 MachOObject::LoadFromBuffer(InputBuffer, &ErrorStr));
469 return Error("unable to load object: '" + ErrorStr + "'");
471 // Get the CPU type information from the header.
472 const macho::Header &Header = Obj->getHeader();
474 // FIXME: Error checking that the loaded object is compatible with
475 // the system we're running on.
476 CPUType = Header.CPUType;
477 CPUSubtype = Header.CPUSubtype;
479 // Validate that the load commands match what we expect.
480 const MachOObject::LoadCommandInfo *SegmentLCI = 0, *SymtabLCI = 0,
482 for (unsigned i = 0; i != Header.NumLoadCommands; ++i) {
483 const MachOObject::LoadCommandInfo &LCI = Obj->getLoadCommandInfo(i);
484 switch (LCI.Command.Type) {
485 case macho::LCT_Segment:
486 case macho::LCT_Segment64:
488 return Error("unexpected input object (multiple segments)");
491 case macho::LCT_Symtab:
493 return Error("unexpected input object (multiple symbol tables)");
496 case macho::LCT_Dysymtab:
498 return Error("unexpected input object (multiple symbol tables)");
502 return Error("unexpected input object (unexpected load command");
507 return Error("no symbol table found in object");
509 return Error("no segments found in object");
511 // Read and register the symbol table data.
512 InMemoryStruct<macho::SymtabLoadCommand> SymtabLC;
513 Obj->ReadSymtabLoadCommand(*SymtabLCI, SymtabLC);
515 return Error("unable to load symbol table load command");
516 Obj->RegisterStringTable(*SymtabLC);
518 // Read the dynamic link-edit information, if present (not present in static
521 InMemoryStruct<macho::DysymtabLoadCommand> DysymtabLC;
522 Obj->ReadDysymtabLoadCommand(*DysymtabLCI, DysymtabLC);
524 return Error("unable to load dynamic link-exit load command");
526 // FIXME: We don't support anything interesting yet.
527 // if (DysymtabLC->LocalSymbolsIndex != 0)
528 // return Error("NOT YET IMPLEMENTED: local symbol entries");
529 // if (DysymtabLC->ExternalSymbolsIndex != 0)
530 // return Error("NOT YET IMPLEMENTED: non-external symbol entries");
531 // if (DysymtabLC->UndefinedSymbolsIndex != SymtabLC->NumSymbolTableEntries)
532 // return Error("NOT YET IMPLEMENTED: undefined symbol entries");
535 // Load the segment load command.
536 if (SegmentLCI->Command.Type == macho::LCT_Segment) {
537 if (loadSegment32(Obj.get(), SegmentLCI, SymtabLC))
540 if (loadSegment64(Obj.get(), SegmentLCI, SymtabLC))
544 // Assign the addresses of the sections from the object so that any
545 // relocations to them get set properly.
546 // FIXME: This is done directly from the client at the moment. We should
547 // default the values to the local storage, at least when the target arch
548 // is the same as the host arch.
553 // Assign an address to a symbol name and resolve all the relocations
554 // associated with it.
555 void RuntimeDyldMachO::reassignSectionAddress(unsigned SectionID,
557 // The address to use for relocation resolution is not
558 // the address of the local section buffer. We must be doing
559 // a remote execution environment of some sort. Re-apply any
560 // relocations referencing this section with the given address.
562 // Addr is a uint64_t because we can't assume the pointer width
563 // of the target is the same as that of the host. Just use a generic
564 // "big enough" type.
566 SectionLoadAddress[SectionID] = Addr;
568 RelocationList &Relocs = Relocations[SectionID];
569 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
570 RelocationEntry &RE = Relocs[i];
571 uint8_t *Target = (uint8_t*)Sections[RE.SectionID].base() + RE.Offset;
572 bool isPCRel = (RE.Data >> 24) & 1;
573 unsigned Type = (RE.Data >> 28) & 0xf;
574 unsigned Size = 1 << ((RE.Data >> 25) & 3);
576 DEBUG(dbgs() << "Resolving relocation at Section #" << RE.SectionID
577 << " + " << RE.Offset << " (" << format("%p", Target) << ")"
578 << " from Section #" << SectionID << " (" << format("%p", Addr) << ")"
579 << "(" << (isPCRel ? "pcrel" : "absolute")
580 << ", type: " << Type << ", Size: " << Size << ", Addend: "
581 << RE.Addend << ").\n");
583 resolveRelocation(Target, Addr, isPCRel, Type, Size, RE.Addend);
587 bool RuntimeDyldMachO::isKnownFormat(const MemoryBuffer *InputBuffer) {
588 StringRef Magic = InputBuffer->getBuffer().slice(0, 4);
589 if (Magic == "\xFE\xED\xFA\xCE") return true;
590 if (Magic == "\xCE\xFA\xED\xFE") return true;
591 if (Magic == "\xFE\xED\xFA\xCF") return true;
592 if (Magic == "\xCF\xFA\xED\xFE") return true;
596 } // end namespace llvm