1 //===- lib/MC/MachObjectWriter.cpp - Mach-O File Writer -------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9
10 #include "llvm/MC/MCMachObjectWriter.h"
11 #include "llvm/ADT/StringMap.h"
12 #include "llvm/ADT/Twine.h"
13 #include "llvm/MC/MCAsmBackend.h"
14 #include "llvm/MC/MCAsmLayout.h"
15 #include "llvm/MC/MCAssembler.h"
16 #include "llvm/MC/MCExpr.h"
17 #include "llvm/MC/MCFixupKindInfo.h"
18 #include "llvm/MC/MCMachOSymbolFlags.h"
19 #include "llvm/MC/MCObjectWriter.h"
20 #include "llvm/MC/MCSectionMachO.h"
21 #include "llvm/MC/MCSymbol.h"
22 #include "llvm/MC/MCValue.h"
23 #include "llvm/Object/MachOFormat.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/ErrorHandling.h"
26 #include <vector>
27 using namespace llvm;
28 using namespace llvm::object;
29
reset()30 void MachObjectWriter::reset() {
31 Relocations.clear();
32 IndirectSymBase.clear();
33 StringTable.clear();
34 LocalSymbolData.clear();
35 ExternalSymbolData.clear();
36 UndefinedSymbolData.clear();
37 MCObjectWriter::reset();
38 }
39
40 bool MachObjectWriter::
doesSymbolRequireExternRelocation(const MCSymbolData * SD)41 doesSymbolRequireExternRelocation(const MCSymbolData *SD) {
42 // Undefined symbols are always extern.
43 if (SD->Symbol->isUndefined())
44 return true;
45
46 // References to weak definitions require external relocation entries; the
47 // definition may not always be the one in the same object file.
48 if (SD->getFlags() & SF_WeakDefinition)
49 return true;
50
51 // Otherwise, we can use an internal relocation.
52 return false;
53 }
54
55 bool MachObjectWriter::
operator <(const MachSymbolData & RHS) const56 MachSymbolData::operator<(const MachSymbolData &RHS) const {
57 return SymbolData->getSymbol().getName() <
58 RHS.SymbolData->getSymbol().getName();
59 }
60
isFixupKindPCRel(const MCAssembler & Asm,unsigned Kind)61 bool MachObjectWriter::isFixupKindPCRel(const MCAssembler &Asm, unsigned Kind) {
62 const MCFixupKindInfo &FKI = Asm.getBackend().getFixupKindInfo(
63 (MCFixupKind) Kind);
64
65 return FKI.Flags & MCFixupKindInfo::FKF_IsPCRel;
66 }
67
getFragmentAddress(const MCFragment * Fragment,const MCAsmLayout & Layout) const68 uint64_t MachObjectWriter::getFragmentAddress(const MCFragment *Fragment,
69 const MCAsmLayout &Layout) const {
70 return getSectionAddress(Fragment->getParent()) +
71 Layout.getFragmentOffset(Fragment);
72 }
73
getSymbolAddress(const MCSymbolData * SD,const MCAsmLayout & Layout) const74 uint64_t MachObjectWriter::getSymbolAddress(const MCSymbolData* SD,
75 const MCAsmLayout &Layout) const {
76 const MCSymbol &S = SD->getSymbol();
77
78 // If this is a variable, then recursively evaluate now.
79 if (S.isVariable()) {
80 if (const MCConstantExpr *C =
81 dyn_cast<const MCConstantExpr>(S.getVariableValue()))
82 return C->getValue();
83
84
85 MCValue Target;
86 if (!S.getVariableValue()->EvaluateAsRelocatable(Target, Layout))
87 report_fatal_error("unable to evaluate offset for variable '" +
88 S.getName() + "'");
89
90 // Verify that any used symbols are defined.
91 if (Target.getSymA() && Target.getSymA()->getSymbol().isUndefined())
92 report_fatal_error("unable to evaluate offset to undefined symbol '" +
93 Target.getSymA()->getSymbol().getName() + "'");
94 if (Target.getSymB() && Target.getSymB()->getSymbol().isUndefined())
95 report_fatal_error("unable to evaluate offset to undefined symbol '" +
96 Target.getSymB()->getSymbol().getName() + "'");
97
98 uint64_t Address = Target.getConstant();
99 if (Target.getSymA())
100 Address += getSymbolAddress(&Layout.getAssembler().getSymbolData(
101 Target.getSymA()->getSymbol()), Layout);
102 if (Target.getSymB())
103 Address += getSymbolAddress(&Layout.getAssembler().getSymbolData(
104 Target.getSymB()->getSymbol()), Layout);
105 return Address;
106 }
107
108 return getSectionAddress(SD->getFragment()->getParent()) +
109 Layout.getSymbolOffset(SD);
110 }
111
getPaddingSize(const MCSectionData * SD,const MCAsmLayout & Layout) const112 uint64_t MachObjectWriter::getPaddingSize(const MCSectionData *SD,
113 const MCAsmLayout &Layout) const {
114 uint64_t EndAddr = getSectionAddress(SD) + Layout.getSectionAddressSize(SD);
115 unsigned Next = SD->getLayoutOrder() + 1;
116 if (Next >= Layout.getSectionOrder().size())
117 return 0;
118
119 const MCSectionData &NextSD = *Layout.getSectionOrder()[Next];
120 if (NextSD.getSection().isVirtualSection())
121 return 0;
122 return OffsetToAlignment(EndAddr, NextSD.getAlignment());
123 }
124
WriteHeader(unsigned NumLoadCommands,unsigned LoadCommandsSize,bool SubsectionsViaSymbols)125 void MachObjectWriter::WriteHeader(unsigned NumLoadCommands,
126 unsigned LoadCommandsSize,
127 bool SubsectionsViaSymbols) {
128 uint32_t Flags = 0;
129
130 if (SubsectionsViaSymbols)
131 Flags |= macho::HF_SubsectionsViaSymbols;
132
133 // struct mach_header (28 bytes) or
134 // struct mach_header_64 (32 bytes)
135
136 uint64_t Start = OS.tell();
137 (void) Start;
138
139 Write32(is64Bit() ? macho::HM_Object64 : macho::HM_Object32);
140
141 Write32(TargetObjectWriter->getCPUType());
142 Write32(TargetObjectWriter->getCPUSubtype());
143
144 Write32(macho::HFT_Object);
145 Write32(NumLoadCommands);
146 Write32(LoadCommandsSize);
147 Write32(Flags);
148 if (is64Bit())
149 Write32(0); // reserved
150
151 assert(OS.tell() - Start ==
152 (is64Bit() ? macho::Header64Size : macho::Header32Size));
153 }
154
155 /// WriteSegmentLoadCommand - Write a segment load command.
156 ///
157 /// \param NumSections The number of sections in this segment.
158 /// \param SectionDataSize The total size of the sections.
WriteSegmentLoadCommand(unsigned NumSections,uint64_t VMSize,uint64_t SectionDataStartOffset,uint64_t SectionDataSize)159 void MachObjectWriter::WriteSegmentLoadCommand(unsigned NumSections,
160 uint64_t VMSize,
161 uint64_t SectionDataStartOffset,
162 uint64_t SectionDataSize) {
163 // struct segment_command (56 bytes) or
164 // struct segment_command_64 (72 bytes)
165
166 uint64_t Start = OS.tell();
167 (void) Start;
168
169 unsigned SegmentLoadCommandSize =
170 is64Bit() ? macho::SegmentLoadCommand64Size:
171 macho::SegmentLoadCommand32Size;
172 Write32(is64Bit() ? macho::LCT_Segment64 : macho::LCT_Segment);
173 Write32(SegmentLoadCommandSize +
174 NumSections * (is64Bit() ? macho::Section64Size :
175 macho::Section32Size));
176
177 WriteBytes("", 16);
178 if (is64Bit()) {
179 Write64(0); // vmaddr
180 Write64(VMSize); // vmsize
181 Write64(SectionDataStartOffset); // file offset
182 Write64(SectionDataSize); // file size
183 } else {
184 Write32(0); // vmaddr
185 Write32(VMSize); // vmsize
186 Write32(SectionDataStartOffset); // file offset
187 Write32(SectionDataSize); // file size
188 }
189 Write32(0x7); // maxprot
190 Write32(0x7); // initprot
191 Write32(NumSections);
192 Write32(0); // flags
193
194 assert(OS.tell() - Start == SegmentLoadCommandSize);
195 }
196
WriteSection(const MCAssembler & Asm,const MCAsmLayout & Layout,const MCSectionData & SD,uint64_t FileOffset,uint64_t RelocationsStart,unsigned NumRelocations)197 void MachObjectWriter::WriteSection(const MCAssembler &Asm,
198 const MCAsmLayout &Layout,
199 const MCSectionData &SD,
200 uint64_t FileOffset,
201 uint64_t RelocationsStart,
202 unsigned NumRelocations) {
203 uint64_t SectionSize = Layout.getSectionAddressSize(&SD);
204
205 // The offset is unused for virtual sections.
206 if (SD.getSection().isVirtualSection()) {
207 assert(Layout.getSectionFileSize(&SD) == 0 && "Invalid file size!");
208 FileOffset = 0;
209 }
210
211 // struct section (68 bytes) or
212 // struct section_64 (80 bytes)
213
214 uint64_t Start = OS.tell();
215 (void) Start;
216
217 const MCSectionMachO &Section = cast<MCSectionMachO>(SD.getSection());
218 WriteBytes(Section.getSectionName(), 16);
219 WriteBytes(Section.getSegmentName(), 16);
220 if (is64Bit()) {
221 Write64(getSectionAddress(&SD)); // address
222 Write64(SectionSize); // size
223 } else {
224 Write32(getSectionAddress(&SD)); // address
225 Write32(SectionSize); // size
226 }
227 Write32(FileOffset);
228
229 unsigned Flags = Section.getTypeAndAttributes();
230 if (SD.hasInstructions())
231 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
232
233 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
234 Write32(Log2_32(SD.getAlignment()));
235 Write32(NumRelocations ? RelocationsStart : 0);
236 Write32(NumRelocations);
237 Write32(Flags);
238 Write32(IndirectSymBase.lookup(&SD)); // reserved1
239 Write32(Section.getStubSize()); // reserved2
240 if (is64Bit())
241 Write32(0); // reserved3
242
243 assert(OS.tell() - Start == (is64Bit() ? macho::Section64Size :
244 macho::Section32Size));
245 }
246
WriteSymtabLoadCommand(uint32_t SymbolOffset,uint32_t NumSymbols,uint32_t StringTableOffset,uint32_t StringTableSize)247 void MachObjectWriter::WriteSymtabLoadCommand(uint32_t SymbolOffset,
248 uint32_t NumSymbols,
249 uint32_t StringTableOffset,
250 uint32_t StringTableSize) {
251 // struct symtab_command (24 bytes)
252
253 uint64_t Start = OS.tell();
254 (void) Start;
255
256 Write32(macho::LCT_Symtab);
257 Write32(macho::SymtabLoadCommandSize);
258 Write32(SymbolOffset);
259 Write32(NumSymbols);
260 Write32(StringTableOffset);
261 Write32(StringTableSize);
262
263 assert(OS.tell() - Start == macho::SymtabLoadCommandSize);
264 }
265
WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,uint32_t NumLocalSymbols,uint32_t FirstExternalSymbol,uint32_t NumExternalSymbols,uint32_t FirstUndefinedSymbol,uint32_t NumUndefinedSymbols,uint32_t IndirectSymbolOffset,uint32_t NumIndirectSymbols)266 void MachObjectWriter::WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
267 uint32_t NumLocalSymbols,
268 uint32_t FirstExternalSymbol,
269 uint32_t NumExternalSymbols,
270 uint32_t FirstUndefinedSymbol,
271 uint32_t NumUndefinedSymbols,
272 uint32_t IndirectSymbolOffset,
273 uint32_t NumIndirectSymbols) {
274 // struct dysymtab_command (80 bytes)
275
276 uint64_t Start = OS.tell();
277 (void) Start;
278
279 Write32(macho::LCT_Dysymtab);
280 Write32(macho::DysymtabLoadCommandSize);
281 Write32(FirstLocalSymbol);
282 Write32(NumLocalSymbols);
283 Write32(FirstExternalSymbol);
284 Write32(NumExternalSymbols);
285 Write32(FirstUndefinedSymbol);
286 Write32(NumUndefinedSymbols);
287 Write32(0); // tocoff
288 Write32(0); // ntoc
289 Write32(0); // modtaboff
290 Write32(0); // nmodtab
291 Write32(0); // extrefsymoff
292 Write32(0); // nextrefsyms
293 Write32(IndirectSymbolOffset);
294 Write32(NumIndirectSymbols);
295 Write32(0); // extreloff
296 Write32(0); // nextrel
297 Write32(0); // locreloff
298 Write32(0); // nlocrel
299
300 assert(OS.tell() - Start == macho::DysymtabLoadCommandSize);
301 }
302
WriteNlist(MachSymbolData & MSD,const MCAsmLayout & Layout)303 void MachObjectWriter::WriteNlist(MachSymbolData &MSD,
304 const MCAsmLayout &Layout) {
305 MCSymbolData &Data = *MSD.SymbolData;
306 const MCSymbol &Symbol = Data.getSymbol();
307 uint8_t Type = 0;
308 uint16_t Flags = Data.getFlags();
309 uint64_t Address = 0;
310
311 // Set the N_TYPE bits. See <mach-o/nlist.h>.
312 //
313 // FIXME: Are the prebound or indirect fields possible here?
314 if (Symbol.isUndefined())
315 Type = macho::STT_Undefined;
316 else if (Symbol.isAbsolute())
317 Type = macho::STT_Absolute;
318 else
319 Type = macho::STT_Section;
320
321 // FIXME: Set STAB bits.
322
323 if (Data.isPrivateExtern())
324 Type |= macho::STF_PrivateExtern;
325
326 // Set external bit.
327 if (Data.isExternal() || Symbol.isUndefined())
328 Type |= macho::STF_External;
329
330 // Compute the symbol address.
331 if (Symbol.isDefined()) {
332 Address = getSymbolAddress(&Data, Layout);
333 } else if (Data.isCommon()) {
334 // Common symbols are encoded with the size in the address
335 // field, and their alignment in the flags.
336 Address = Data.getCommonSize();
337
338 // Common alignment is packed into the 'desc' bits.
339 if (unsigned Align = Data.getCommonAlignment()) {
340 unsigned Log2Size = Log2_32(Align);
341 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
342 if (Log2Size > 15)
343 report_fatal_error("invalid 'common' alignment '" +
344 Twine(Align) + "'");
345 // FIXME: Keep this mask with the SymbolFlags enumeration.
346 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
347 }
348 }
349
350 // struct nlist (12 bytes)
351
352 Write32(MSD.StringIndex);
353 Write8(Type);
354 Write8(MSD.SectionIndex);
355
356 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
357 // value.
358 Write16(Flags);
359 if (is64Bit())
360 Write64(Address);
361 else
362 Write32(Address);
363 }
364
WriteLinkeditLoadCommand(uint32_t Type,uint32_t DataOffset,uint32_t DataSize)365 void MachObjectWriter::WriteLinkeditLoadCommand(uint32_t Type,
366 uint32_t DataOffset,
367 uint32_t DataSize) {
368 uint64_t Start = OS.tell();
369 (void) Start;
370
371 Write32(Type);
372 Write32(macho::LinkeditLoadCommandSize);
373 Write32(DataOffset);
374 Write32(DataSize);
375
376 assert(OS.tell() - Start == macho::LinkeditLoadCommandSize);
377 }
378
ComputeLinkerOptionsLoadCommandSize(const std::vector<std::string> & Options,bool is64Bit)379 static unsigned ComputeLinkerOptionsLoadCommandSize(
380 const std::vector<std::string> &Options, bool is64Bit)
381 {
382 unsigned Size = sizeof(macho::LinkerOptionsLoadCommand);
383 for (unsigned i = 0, e = Options.size(); i != e; ++i)
384 Size += Options[i].size() + 1;
385 return RoundUpToAlignment(Size, is64Bit ? 8 : 4);
386 }
387
WriteLinkerOptionsLoadCommand(const std::vector<std::string> & Options)388 void MachObjectWriter::WriteLinkerOptionsLoadCommand(
389 const std::vector<std::string> &Options)
390 {
391 unsigned Size = ComputeLinkerOptionsLoadCommandSize(Options, is64Bit());
392 uint64_t Start = OS.tell();
393 (void) Start;
394
395 Write32(macho::LCT_LinkerOptions);
396 Write32(Size);
397 Write32(Options.size());
398 uint64_t BytesWritten = sizeof(macho::LinkerOptionsLoadCommand);
399 for (unsigned i = 0, e = Options.size(); i != e; ++i) {
400 // Write each string, including the null byte.
401 const std::string &Option = Options[i];
402 WriteBytes(Option.c_str(), Option.size() + 1);
403 BytesWritten += Option.size() + 1;
404 }
405
406 // Pad to a multiple of the pointer size.
407 WriteBytes("", OffsetToAlignment(BytesWritten, is64Bit() ? 8 : 4));
408
409 assert(OS.tell() - Start == Size);
410 }
411
412
RecordRelocation(const MCAssembler & Asm,const MCAsmLayout & Layout,const MCFragment * Fragment,const MCFixup & Fixup,MCValue Target,uint64_t & FixedValue)413 void MachObjectWriter::RecordRelocation(const MCAssembler &Asm,
414 const MCAsmLayout &Layout,
415 const MCFragment *Fragment,
416 const MCFixup &Fixup,
417 MCValue Target,
418 uint64_t &FixedValue) {
419 TargetObjectWriter->RecordRelocation(this, Asm, Layout, Fragment, Fixup,
420 Target, FixedValue);
421 }
422
BindIndirectSymbols(MCAssembler & Asm)423 void MachObjectWriter::BindIndirectSymbols(MCAssembler &Asm) {
424 // This is the point where 'as' creates actual symbols for indirect symbols
425 // (in the following two passes). It would be easier for us to do this sooner
426 // when we see the attribute, but that makes getting the order in the symbol
427 // table much more complicated than it is worth.
428 //
429 // FIXME: Revisit this when the dust settles.
430
431 // Bind non lazy symbol pointers first.
432 unsigned IndirectIndex = 0;
433 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
434 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
435 const MCSectionMachO &Section =
436 cast<MCSectionMachO>(it->SectionData->getSection());
437
438 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
439 continue;
440
441 // Initialize the section indirect symbol base, if necessary.
442 IndirectSymBase.insert(std::make_pair(it->SectionData, IndirectIndex));
443
444 Asm.getOrCreateSymbolData(*it->Symbol);
445 }
446
447 // Then lazy symbol pointers and symbol stubs.
448 IndirectIndex = 0;
449 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
450 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
451 const MCSectionMachO &Section =
452 cast<MCSectionMachO>(it->SectionData->getSection());
453
454 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
455 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
456 continue;
457
458 // Initialize the section indirect symbol base, if necessary.
459 IndirectSymBase.insert(std::make_pair(it->SectionData, IndirectIndex));
460
461 // Set the symbol type to undefined lazy, but only on construction.
462 //
463 // FIXME: Do not hardcode.
464 bool Created;
465 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
466 if (Created)
467 Entry.setFlags(Entry.getFlags() | 0x0001);
468 }
469 }
470
471 /// ComputeSymbolTable - Compute the symbol table data
472 ///
473 /// \param StringTable [out] - The string table data.
474 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
475 /// string table.
476 void MachObjectWriter::
ComputeSymbolTable(MCAssembler & Asm,SmallString<256> & StringTable,std::vector<MachSymbolData> & LocalSymbolData,std::vector<MachSymbolData> & ExternalSymbolData,std::vector<MachSymbolData> & UndefinedSymbolData)477 ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
478 std::vector<MachSymbolData> &LocalSymbolData,
479 std::vector<MachSymbolData> &ExternalSymbolData,
480 std::vector<MachSymbolData> &UndefinedSymbolData) {
481 // Build section lookup table.
482 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
483 unsigned Index = 1;
484 for (MCAssembler::iterator it = Asm.begin(),
485 ie = Asm.end(); it != ie; ++it, ++Index)
486 SectionIndexMap[&it->getSection()] = Index;
487 assert(Index <= 256 && "Too many sections!");
488
489 // Index 0 is always the empty string.
490 StringMap<uint64_t> StringIndexMap;
491 StringTable += '\x00';
492
493 // Build the symbol arrays and the string table, but only for non-local
494 // symbols.
495 //
496 // The particular order that we collect the symbols and create the string
497 // table, then sort the symbols is chosen to match 'as'. Even though it
498 // doesn't matter for correctness, this is important for letting us diff .o
499 // files.
500 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
501 ie = Asm.symbol_end(); it != ie; ++it) {
502 const MCSymbol &Symbol = it->getSymbol();
503
504 // Ignore non-linker visible symbols.
505 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
506 continue;
507
508 if (!it->isExternal() && !Symbol.isUndefined())
509 continue;
510
511 uint64_t &Entry = StringIndexMap[Symbol.getName()];
512 if (!Entry) {
513 Entry = StringTable.size();
514 StringTable += Symbol.getName();
515 StringTable += '\x00';
516 }
517
518 MachSymbolData MSD;
519 MSD.SymbolData = it;
520 MSD.StringIndex = Entry;
521
522 if (Symbol.isUndefined()) {
523 MSD.SectionIndex = 0;
524 UndefinedSymbolData.push_back(MSD);
525 } else if (Symbol.isAbsolute()) {
526 MSD.SectionIndex = 0;
527 ExternalSymbolData.push_back(MSD);
528 } else {
529 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
530 assert(MSD.SectionIndex && "Invalid section index!");
531 ExternalSymbolData.push_back(MSD);
532 }
533 }
534
535 // Now add the data for local symbols.
536 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
537 ie = Asm.symbol_end(); it != ie; ++it) {
538 const MCSymbol &Symbol = it->getSymbol();
539
540 // Ignore non-linker visible symbols.
541 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
542 continue;
543
544 if (it->isExternal() || Symbol.isUndefined())
545 continue;
546
547 uint64_t &Entry = StringIndexMap[Symbol.getName()];
548 if (!Entry) {
549 Entry = StringTable.size();
550 StringTable += Symbol.getName();
551 StringTable += '\x00';
552 }
553
554 MachSymbolData MSD;
555 MSD.SymbolData = it;
556 MSD.StringIndex = Entry;
557
558 if (Symbol.isAbsolute()) {
559 MSD.SectionIndex = 0;
560 LocalSymbolData.push_back(MSD);
561 } else {
562 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
563 assert(MSD.SectionIndex && "Invalid section index!");
564 LocalSymbolData.push_back(MSD);
565 }
566 }
567
568 // External and undefined symbols are required to be in lexicographic order.
569 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
570 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
571
572 // Set the symbol indices.
573 Index = 0;
574 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
575 LocalSymbolData[i].SymbolData->setIndex(Index++);
576 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
577 ExternalSymbolData[i].SymbolData->setIndex(Index++);
578 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
579 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
580
581 // The string table is padded to a multiple of 4.
582 while (StringTable.size() % 4)
583 StringTable += '\x00';
584 }
585
computeSectionAddresses(const MCAssembler & Asm,const MCAsmLayout & Layout)586 void MachObjectWriter::computeSectionAddresses(const MCAssembler &Asm,
587 const MCAsmLayout &Layout) {
588 uint64_t StartAddress = 0;
589 const SmallVectorImpl<MCSectionData*> &Order = Layout.getSectionOrder();
590 for (int i = 0, n = Order.size(); i != n ; ++i) {
591 const MCSectionData *SD = Order[i];
592 StartAddress = RoundUpToAlignment(StartAddress, SD->getAlignment());
593 SectionAddress[SD] = StartAddress;
594 StartAddress += Layout.getSectionAddressSize(SD);
595
596 // Explicitly pad the section to match the alignment requirements of the
597 // following one. This is for 'gas' compatibility, it shouldn't
598 /// strictly be necessary.
599 StartAddress += getPaddingSize(SD, Layout);
600 }
601 }
602
markAbsoluteVariableSymbols(MCAssembler & Asm,const MCAsmLayout & Layout)603 void MachObjectWriter::markAbsoluteVariableSymbols(MCAssembler &Asm,
604 const MCAsmLayout &Layout) {
605 for (MCAssembler::symbol_iterator i = Asm.symbol_begin(),
606 e = Asm.symbol_end();
607 i != e; ++i) {
608 MCSymbolData &SD = *i;
609 if (!SD.getSymbol().isVariable())
610 continue;
611
612 // Is the variable is a symbol difference (SA - SB + C) expression,
613 // and neither symbol is external, mark the variable as absolute.
614 const MCExpr *Expr = SD.getSymbol().getVariableValue();
615 MCValue Value;
616 if (Expr->EvaluateAsRelocatable(Value, Layout)) {
617 if (Value.getSymA() && Value.getSymB())
618 const_cast<MCSymbol*>(&SD.getSymbol())->setAbsolute();
619 }
620 }
621 }
622
ExecutePostLayoutBinding(MCAssembler & Asm,const MCAsmLayout & Layout)623 void MachObjectWriter::ExecutePostLayoutBinding(MCAssembler &Asm,
624 const MCAsmLayout &Layout) {
625 computeSectionAddresses(Asm, Layout);
626
627 // Create symbol data for any indirect symbols.
628 BindIndirectSymbols(Asm);
629
630 // Mark symbol difference expressions in variables (from .set or = directives)
631 // as absolute.
632 markAbsoluteVariableSymbols(Asm, Layout);
633
634 // Compute symbol table information and bind symbol indices.
635 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
636 UndefinedSymbolData);
637 }
638
639 bool MachObjectWriter::
IsSymbolRefDifferenceFullyResolvedImpl(const MCAssembler & Asm,const MCSymbolData & DataA,const MCFragment & FB,bool InSet,bool IsPCRel) const640 IsSymbolRefDifferenceFullyResolvedImpl(const MCAssembler &Asm,
641 const MCSymbolData &DataA,
642 const MCFragment &FB,
643 bool InSet,
644 bool IsPCRel) const {
645 if (InSet)
646 return true;
647
648 // The effective address is
649 // addr(atom(A)) + offset(A)
650 // - addr(atom(B)) - offset(B)
651 // and the offsets are not relocatable, so the fixup is fully resolved when
652 // addr(atom(A)) - addr(atom(B)) == 0.
653 const MCSymbolData *A_Base = 0, *B_Base = 0;
654
655 const MCSymbol &SA = DataA.getSymbol().AliasedSymbol();
656 const MCSection &SecA = SA.getSection();
657 const MCSection &SecB = FB.getParent()->getSection();
658
659 if (IsPCRel) {
660 // The simple (Darwin, except on x86_64) way of dealing with this was to
661 // assume that any reference to a temporary symbol *must* be a temporary
662 // symbol in the same atom, unless the sections differ. Therefore, any PCrel
663 // relocation to a temporary symbol (in the same section) is fully
664 // resolved. This also works in conjunction with absolutized .set, which
665 // requires the compiler to use .set to absolutize the differences between
666 // symbols which the compiler knows to be assembly time constants, so we
667 // don't need to worry about considering symbol differences fully resolved.
668 //
669 // If the file isn't using sub-sections-via-symbols, we can make the
670 // same assumptions about any symbol that we normally make about
671 // assembler locals.
672
673 if (!Asm.getBackend().hasReliableSymbolDifference()) {
674 if (!SA.isInSection() || &SecA != &SecB ||
675 (!SA.isTemporary() &&
676 FB.getAtom() != Asm.getSymbolData(SA).getFragment()->getAtom() &&
677 Asm.getSubsectionsViaSymbols()))
678 return false;
679 return true;
680 }
681 // For Darwin x86_64, there is one special case when the reference IsPCRel.
682 // If the fragment with the reference does not have a base symbol but meets
683 // the simple way of dealing with this, in that it is a temporary symbol in
684 // the same atom then it is assumed to be fully resolved. This is needed so
685 // a relocation entry is not created and so the static linker does not
686 // mess up the reference later.
687 else if(!FB.getAtom() &&
688 SA.isTemporary() && SA.isInSection() && &SecA == &SecB){
689 return true;
690 }
691 } else {
692 if (!TargetObjectWriter->useAggressiveSymbolFolding())
693 return false;
694 }
695
696 const MCFragment *FA = Asm.getSymbolData(SA).getFragment();
697
698 // Bail if the symbol has no fragment.
699 if (!FA)
700 return false;
701
702 A_Base = FA->getAtom();
703 if (!A_Base)
704 return false;
705
706 B_Base = FB.getAtom();
707 if (!B_Base)
708 return false;
709
710 // If the atoms are the same, they are guaranteed to have the same address.
711 if (A_Base == B_Base)
712 return true;
713
714 // Otherwise, we can't prove this is fully resolved.
715 return false;
716 }
717
WriteObject(MCAssembler & Asm,const MCAsmLayout & Layout)718 void MachObjectWriter::WriteObject(MCAssembler &Asm,
719 const MCAsmLayout &Layout) {
720 unsigned NumSections = Asm.size();
721
722 // The section data starts after the header, the segment load command (and
723 // section headers) and the symbol table.
724 unsigned NumLoadCommands = 1;
725 uint64_t LoadCommandsSize = is64Bit() ?
726 macho::SegmentLoadCommand64Size + NumSections * macho::Section64Size :
727 macho::SegmentLoadCommand32Size + NumSections * macho::Section32Size;
728
729 // Add the data-in-code load command size, if used.
730 unsigned NumDataRegions = Asm.getDataRegions().size();
731 if (NumDataRegions) {
732 ++NumLoadCommands;
733 LoadCommandsSize += macho::LinkeditLoadCommandSize;
734 }
735
736 // Add the symbol table load command sizes, if used.
737 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
738 UndefinedSymbolData.size();
739 if (NumSymbols) {
740 NumLoadCommands += 2;
741 LoadCommandsSize += (macho::SymtabLoadCommandSize +
742 macho::DysymtabLoadCommandSize);
743 }
744
745 // Add the linker option load commands sizes.
746 const std::vector<std::vector<std::string> > &LinkerOptions =
747 Asm.getLinkerOptions();
748 for (unsigned i = 0, e = LinkerOptions.size(); i != e; ++i) {
749 ++NumLoadCommands;
750 LoadCommandsSize += ComputeLinkerOptionsLoadCommandSize(LinkerOptions[i],
751 is64Bit());
752 }
753
754 // Compute the total size of the section data, as well as its file size and vm
755 // size.
756 uint64_t SectionDataStart = (is64Bit() ? macho::Header64Size :
757 macho::Header32Size) + LoadCommandsSize;
758 uint64_t SectionDataSize = 0;
759 uint64_t SectionDataFileSize = 0;
760 uint64_t VMSize = 0;
761 for (MCAssembler::const_iterator it = Asm.begin(),
762 ie = Asm.end(); it != ie; ++it) {
763 const MCSectionData &SD = *it;
764 uint64_t Address = getSectionAddress(&SD);
765 uint64_t Size = Layout.getSectionAddressSize(&SD);
766 uint64_t FileSize = Layout.getSectionFileSize(&SD);
767 FileSize += getPaddingSize(&SD, Layout);
768
769 VMSize = std::max(VMSize, Address + Size);
770
771 if (SD.getSection().isVirtualSection())
772 continue;
773
774 SectionDataSize = std::max(SectionDataSize, Address + Size);
775 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
776 }
777
778 // The section data is padded to 4 bytes.
779 //
780 // FIXME: Is this machine dependent?
781 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
782 SectionDataFileSize += SectionDataPadding;
783
784 // Write the prolog, starting with the header and load command...
785 WriteHeader(NumLoadCommands, LoadCommandsSize,
786 Asm.getSubsectionsViaSymbols());
787 WriteSegmentLoadCommand(NumSections, VMSize,
788 SectionDataStart, SectionDataSize);
789
790 // ... and then the section headers.
791 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
792 for (MCAssembler::const_iterator it = Asm.begin(),
793 ie = Asm.end(); it != ie; ++it) {
794 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
795 unsigned NumRelocs = Relocs.size();
796 uint64_t SectionStart = SectionDataStart + getSectionAddress(it);
797 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
798 RelocTableEnd += NumRelocs * macho::RelocationInfoSize;
799 }
800
801 // Write the data-in-code load command, if used.
802 uint64_t DataInCodeTableEnd = RelocTableEnd + NumDataRegions * 8;
803 if (NumDataRegions) {
804 uint64_t DataRegionsOffset = RelocTableEnd;
805 uint64_t DataRegionsSize = NumDataRegions * 8;
806 WriteLinkeditLoadCommand(macho::LCT_DataInCode, DataRegionsOffset,
807 DataRegionsSize);
808 }
809
810 // Write the symbol table load command, if used.
811 if (NumSymbols) {
812 unsigned FirstLocalSymbol = 0;
813 unsigned NumLocalSymbols = LocalSymbolData.size();
814 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
815 unsigned NumExternalSymbols = ExternalSymbolData.size();
816 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
817 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
818 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
819 unsigned NumSymTabSymbols =
820 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
821 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
822 uint64_t IndirectSymbolOffset = 0;
823
824 // If used, the indirect symbols are written after the section data.
825 if (NumIndirectSymbols)
826 IndirectSymbolOffset = DataInCodeTableEnd;
827
828 // The symbol table is written after the indirect symbol data.
829 uint64_t SymbolTableOffset = DataInCodeTableEnd + IndirectSymbolSize;
830
831 // The string table is written after symbol table.
832 uint64_t StringTableOffset =
833 SymbolTableOffset + NumSymTabSymbols * (is64Bit() ? macho::Nlist64Size :
834 macho::Nlist32Size);
835 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
836 StringTableOffset, StringTable.size());
837
838 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
839 FirstExternalSymbol, NumExternalSymbols,
840 FirstUndefinedSymbol, NumUndefinedSymbols,
841 IndirectSymbolOffset, NumIndirectSymbols);
842 }
843
844 // Write the linker options load commands.
845 for (unsigned i = 0, e = LinkerOptions.size(); i != e; ++i) {
846 WriteLinkerOptionsLoadCommand(LinkerOptions[i]);
847 }
848
849 // Write the actual section data.
850 for (MCAssembler::const_iterator it = Asm.begin(),
851 ie = Asm.end(); it != ie; ++it) {
852 Asm.writeSectionData(it, Layout);
853
854 uint64_t Pad = getPaddingSize(it, Layout);
855 for (unsigned int i = 0; i < Pad; ++i)
856 Write8(0);
857 }
858
859 // Write the extra padding.
860 WriteZeros(SectionDataPadding);
861
862 // Write the relocation entries.
863 for (MCAssembler::const_iterator it = Asm.begin(),
864 ie = Asm.end(); it != ie; ++it) {
865 // Write the section relocation entries, in reverse order to match 'as'
866 // (approximately, the exact algorithm is more complicated than this).
867 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
868 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
869 Write32(Relocs[e - i - 1].Word0);
870 Write32(Relocs[e - i - 1].Word1);
871 }
872 }
873
874 // Write out the data-in-code region payload, if there is one.
875 for (MCAssembler::const_data_region_iterator
876 it = Asm.data_region_begin(), ie = Asm.data_region_end();
877 it != ie; ++it) {
878 const DataRegionData *Data = &(*it);
879 uint64_t Start =
880 getSymbolAddress(&Layout.getAssembler().getSymbolData(*Data->Start),
881 Layout);
882 uint64_t End =
883 getSymbolAddress(&Layout.getAssembler().getSymbolData(*Data->End),
884 Layout);
885 DEBUG(dbgs() << "data in code region-- kind: " << Data->Kind
886 << " start: " << Start << "(" << Data->Start->getName() << ")"
887 << " end: " << End << "(" << Data->End->getName() << ")"
888 << " size: " << End - Start
889 << "\n");
890 Write32(Start);
891 Write16(End - Start);
892 Write16(Data->Kind);
893 }
894
895 // Write the symbol table data, if used.
896 if (NumSymbols) {
897 // Write the indirect symbol entries.
898 for (MCAssembler::const_indirect_symbol_iterator
899 it = Asm.indirect_symbol_begin(),
900 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
901 // Indirect symbols in the non lazy symbol pointer section have some
902 // special handling.
903 const MCSectionMachO &Section =
904 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
905 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
906 // If this symbol is defined and internal, mark it as such.
907 if (it->Symbol->isDefined() &&
908 !Asm.getSymbolData(*it->Symbol).isExternal()) {
909 uint32_t Flags = macho::ISF_Local;
910 if (it->Symbol->isAbsolute())
911 Flags |= macho::ISF_Absolute;
912 Write32(Flags);
913 continue;
914 }
915 }
916
917 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
918 }
919
920 // FIXME: Check that offsets match computed ones.
921
922 // Write the symbol table entries.
923 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
924 WriteNlist(LocalSymbolData[i], Layout);
925 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
926 WriteNlist(ExternalSymbolData[i], Layout);
927 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
928 WriteNlist(UndefinedSymbolData[i], Layout);
929
930 // Write the string table.
931 OS << StringTable.str();
932 }
933 }
934
createMachObjectWriter(MCMachObjectTargetWriter * MOTW,raw_ostream & OS,bool IsLittleEndian)935 MCObjectWriter *llvm::createMachObjectWriter(MCMachObjectTargetWriter *MOTW,
936 raw_ostream &OS,
937 bool IsLittleEndian) {
938 return new MachObjectWriter(MOTW, OS, IsLittleEndian);
939 }
940