1 //===- lib/MC/MachObjectWriter.cpp - Mach-O File Writer -------------------===//
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 #include "llvm/MC/MCMachObjectWriter.h"
11 #include "llvm/ADT/OwningPtr.h"
12 #include "llvm/ADT/StringMap.h"
13 #include "llvm/ADT/Twine.h"
14 #include "llvm/MC/MCAssembler.h"
15 #include "llvm/MC/MCAsmLayout.h"
16 #include "llvm/MC/MCExpr.h"
17 #include "llvm/MC/MCObjectWriter.h"
18 #include "llvm/MC/MCSectionMachO.h"
19 #include "llvm/MC/MCSymbol.h"
20 #include "llvm/MC/MCMachOSymbolFlags.h"
21 #include "llvm/MC/MCValue.h"
22 #include "llvm/Object/MachOFormat.h"
23 #include "llvm/Support/ErrorHandling.h"
24 #include "llvm/Target/TargetAsmBackend.h"
27 #include "../Target/ARM/ARMFixupKinds.h"
28 #include "../Target/X86/X86FixupKinds.h"
32 using namespace llvm::object;
34 // FIXME: this has been copied from (or to) X86AsmBackend.cpp
35 static unsigned getFixupKindLog2Size(unsigned Kind) {
38 llvm_unreachable("invalid fixup kind!");
40 case FK_Data_1: return 0;
42 case FK_Data_2: return 1;
44 // FIXME: Remove these!!!
45 case X86::reloc_riprel_4byte:
46 case X86::reloc_riprel_4byte_movq_load:
47 case X86::reloc_signed_4byte:
48 case FK_Data_4: return 2;
49 case FK_Data_8: return 3;
53 static bool doesSymbolRequireExternRelocation(MCSymbolData *SD) {
54 // Undefined symbols are always extern.
55 if (SD->Symbol->isUndefined())
58 // References to weak definitions require external relocation entries; the
59 // definition may not always be the one in the same object file.
60 if (SD->getFlags() & SF_WeakDefinition)
63 // Otherwise, we can use an internal relocation.
69 class MachObjectWriter : public MCObjectWriter {
70 /// MachSymbolData - Helper struct for containing some precomputed information
72 struct MachSymbolData {
73 MCSymbolData *SymbolData;
77 // Support lexicographic sorting.
78 bool operator<(const MachSymbolData &RHS) const {
79 return SymbolData->getSymbol().getName() <
80 RHS.SymbolData->getSymbol().getName();
84 /// The target specific Mach-O writer instance.
85 llvm::OwningPtr<MCMachObjectTargetWriter> TargetObjectWriter;
87 /// @name Relocation Data
90 llvm::DenseMap<const MCSectionData*,
91 std::vector<macho::RelocationEntry> > Relocations;
92 llvm::DenseMap<const MCSectionData*, unsigned> IndirectSymBase;
95 /// @name Symbol Table Data
98 SmallString<256> StringTable;
99 std::vector<MachSymbolData> LocalSymbolData;
100 std::vector<MachSymbolData> ExternalSymbolData;
101 std::vector<MachSymbolData> UndefinedSymbolData;
106 /// @name Utility Methods
109 bool isFixupKindPCRel(const MCAssembler &Asm, unsigned Kind) {
110 const MCFixupKindInfo &FKI = Asm.getBackend().getFixupKindInfo(
113 return FKI.Flags & MCFixupKindInfo::FKF_IsPCRel;
118 SectionAddrMap SectionAddress;
119 uint64_t getSectionAddress(const MCSectionData* SD) const {
120 return SectionAddress.lookup(SD);
122 uint64_t getSymbolAddress(const MCSymbolData* SD,
123 const MCAsmLayout &Layout) const;
125 uint64_t getFragmentAddress(const MCFragment *Fragment,
126 const MCAsmLayout &Layout) const {
127 return getSectionAddress(Fragment->getParent()) +
128 Layout.getFragmentOffset(Fragment);
131 uint64_t getPaddingSize(const MCSectionData *SD,
132 const MCAsmLayout &Layout) const;
134 MachObjectWriter(MCMachObjectTargetWriter *MOTW, raw_ostream &_OS,
135 bool _IsLittleEndian)
136 : MCObjectWriter(_OS, _IsLittleEndian), TargetObjectWriter(MOTW) {
139 /// @name Target Writer Proxy Accessors
142 bool is64Bit() const { return TargetObjectWriter->is64Bit(); }
144 uint32_t CPUType = TargetObjectWriter->getCPUType() & ~mach::CTFM_ArchMask;
145 return CPUType == mach::CTM_ARM;
150 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
151 bool SubsectionsViaSymbols);
153 /// WriteSegmentLoadCommand - Write a segment load command.
155 /// \arg NumSections - The number of sections in this segment.
156 /// \arg SectionDataSize - The total size of the sections.
157 void WriteSegmentLoadCommand(unsigned NumSections,
159 uint64_t SectionDataStartOffset,
160 uint64_t SectionDataSize);
162 void WriteSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
163 const MCSectionData &SD, uint64_t FileOffset,
164 uint64_t RelocationsStart, unsigned NumRelocations);
166 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
167 uint32_t StringTableOffset,
168 uint32_t StringTableSize);
170 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
171 uint32_t NumLocalSymbols,
172 uint32_t FirstExternalSymbol,
173 uint32_t NumExternalSymbols,
174 uint32_t FirstUndefinedSymbol,
175 uint32_t NumUndefinedSymbols,
176 uint32_t IndirectSymbolOffset,
177 uint32_t NumIndirectSymbols);
179 void WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout);
181 // FIXME: We really need to improve the relocation validation. Basically, we
182 // want to implement a separate computation which evaluates the relocation
183 // entry as the linker would, and verifies that the resultant fixup value is
184 // exactly what the encoder wanted. This will catch several classes of
187 // - Relocation entry bugs, the two algorithms are unlikely to have the same
190 // - Relaxation issues, where we forget to relax something.
192 // - Input errors, where something cannot be correctly encoded. 'as' allows
193 // these through in many cases.
195 static bool isFixupKindRIPRel(unsigned Kind) {
196 return Kind == X86::reloc_riprel_4byte ||
197 Kind == X86::reloc_riprel_4byte_movq_load;
199 void RecordX86_64Relocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
200 const MCFragment *Fragment,
201 const MCFixup &Fixup, MCValue Target,
202 uint64_t &FixedValue);
204 void RecordScatteredRelocation(const MCAssembler &Asm,
205 const MCAsmLayout &Layout,
206 const MCFragment *Fragment,
207 const MCFixup &Fixup, MCValue Target,
209 uint64_t &FixedValue);
211 void RecordARMScatteredRelocation(const MCAssembler &Asm,
212 const MCAsmLayout &Layout,
213 const MCFragment *Fragment,
214 const MCFixup &Fixup, MCValue Target,
216 uint64_t &FixedValue);
218 void RecordARMMovwMovtRelocation(const MCAssembler &Asm,
219 const MCAsmLayout &Layout,
220 const MCFragment *Fragment,
221 const MCFixup &Fixup, MCValue Target,
222 uint64_t &FixedValue);
224 void RecordTLVPRelocation(const MCAssembler &Asm,
225 const MCAsmLayout &Layout,
226 const MCFragment *Fragment,
227 const MCFixup &Fixup, MCValue Target,
228 uint64_t &FixedValue);
230 static bool getARMFixupKindMachOInfo(unsigned Kind, unsigned &RelocType,
233 void RecordARMRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
234 const MCFragment *Fragment, const MCFixup &Fixup,
235 MCValue Target, uint64_t &FixedValue);
237 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
238 const MCFragment *Fragment, const MCFixup &Fixup,
239 MCValue Target, uint64_t &FixedValue);
241 void BindIndirectSymbols(MCAssembler &Asm);
243 /// ComputeSymbolTable - Compute the symbol table data
245 /// \param StringTable [out] - The string table data.
246 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
248 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
249 std::vector<MachSymbolData> &LocalSymbolData,
250 std::vector<MachSymbolData> &ExternalSymbolData,
251 std::vector<MachSymbolData> &UndefinedSymbolData);
253 void computeSectionAddresses(const MCAssembler &Asm,
254 const MCAsmLayout &Layout);
256 void ExecutePostLayoutBinding(MCAssembler &Asm, const MCAsmLayout &Layout);
258 virtual bool IsSymbolRefDifferenceFullyResolvedImpl(const MCAssembler &Asm,
259 const MCSymbolData &DataA,
260 const MCFragment &FB,
264 void WriteObject(MCAssembler &Asm, const MCAsmLayout &Layout);
267 } // end anonymous namespace
269 uint64_t MachObjectWriter::getSymbolAddress(const MCSymbolData* SD,
270 const MCAsmLayout &Layout) const {
271 const MCSymbol &S = SD->getSymbol();
273 // If this is a variable, then recursively evaluate now.
274 if (S.isVariable()) {
276 if (!S.getVariableValue()->EvaluateAsRelocatable(Target, Layout))
277 report_fatal_error("unable to evaluate offset for variable '" +
280 // Verify that any used symbols are defined.
281 if (Target.getSymA() && Target.getSymA()->getSymbol().isUndefined())
282 report_fatal_error("unable to evaluate offset to undefined symbol '" +
283 Target.getSymA()->getSymbol().getName() + "'");
284 if (Target.getSymB() && Target.getSymB()->getSymbol().isUndefined())
285 report_fatal_error("unable to evaluate offset to undefined symbol '" +
286 Target.getSymB()->getSymbol().getName() + "'");
288 uint64_t Address = Target.getConstant();
289 if (Target.getSymA())
290 Address += getSymbolAddress(&Layout.getAssembler().getSymbolData(
291 Target.getSymA()->getSymbol()), Layout);
292 if (Target.getSymB())
293 Address += getSymbolAddress(&Layout.getAssembler().getSymbolData(
294 Target.getSymB()->getSymbol()), Layout);
298 return getSectionAddress(SD->getFragment()->getParent()) +
299 Layout.getSymbolOffset(SD);
302 uint64_t MachObjectWriter::getPaddingSize(const MCSectionData *SD,
303 const MCAsmLayout &Layout) const {
304 uint64_t EndAddr = getSectionAddress(SD) + Layout.getSectionAddressSize(SD);
305 unsigned Next = SD->getLayoutOrder() + 1;
306 if (Next >= Layout.getSectionOrder().size())
309 const MCSectionData &NextSD = *Layout.getSectionOrder()[Next];
310 if (NextSD.getSection().isVirtualSection())
312 return OffsetToAlignment(EndAddr, NextSD.getAlignment());
315 void MachObjectWriter::WriteHeader(unsigned NumLoadCommands,
316 unsigned LoadCommandsSize,
317 bool SubsectionsViaSymbols) {
320 if (SubsectionsViaSymbols)
321 Flags |= macho::HF_SubsectionsViaSymbols;
323 // struct mach_header (28 bytes) or
324 // struct mach_header_64 (32 bytes)
326 uint64_t Start = OS.tell();
329 Write32(is64Bit() ? macho::HM_Object64 : macho::HM_Object32);
331 Write32(TargetObjectWriter->getCPUType());
332 Write32(TargetObjectWriter->getCPUSubtype());
334 Write32(macho::HFT_Object);
335 Write32(NumLoadCommands);
336 Write32(LoadCommandsSize);
339 Write32(0); // reserved
341 assert(OS.tell() - Start ==
342 (is64Bit() ? macho::Header64Size : macho::Header32Size));
345 /// WriteSegmentLoadCommand - Write a segment load command.
347 /// \arg NumSections - The number of sections in this segment.
348 /// \arg SectionDataSize - The total size of the sections.
349 void MachObjectWriter::WriteSegmentLoadCommand(unsigned NumSections,
351 uint64_t SectionDataStartOffset,
352 uint64_t SectionDataSize) {
353 // struct segment_command (56 bytes) or
354 // struct segment_command_64 (72 bytes)
356 uint64_t Start = OS.tell();
359 unsigned SegmentLoadCommandSize =
360 is64Bit() ? macho::SegmentLoadCommand64Size:
361 macho::SegmentLoadCommand32Size;
362 Write32(is64Bit() ? macho::LCT_Segment64 : macho::LCT_Segment);
363 Write32(SegmentLoadCommandSize +
364 NumSections * (is64Bit() ? macho::Section64Size :
365 macho::Section32Size));
369 Write64(0); // vmaddr
370 Write64(VMSize); // vmsize
371 Write64(SectionDataStartOffset); // file offset
372 Write64(SectionDataSize); // file size
374 Write32(0); // vmaddr
375 Write32(VMSize); // vmsize
376 Write32(SectionDataStartOffset); // file offset
377 Write32(SectionDataSize); // file size
379 Write32(0x7); // maxprot
380 Write32(0x7); // initprot
381 Write32(NumSections);
384 assert(OS.tell() - Start == SegmentLoadCommandSize);
387 void MachObjectWriter::WriteSection(const MCAssembler &Asm,
388 const MCAsmLayout &Layout,
389 const MCSectionData &SD,
391 uint64_t RelocationsStart,
392 unsigned NumRelocations) {
393 uint64_t SectionSize = Layout.getSectionAddressSize(&SD);
395 // The offset is unused for virtual sections.
396 if (SD.getSection().isVirtualSection()) {
397 assert(Layout.getSectionFileSize(&SD) == 0 && "Invalid file size!");
401 // struct section (68 bytes) or
402 // struct section_64 (80 bytes)
404 uint64_t Start = OS.tell();
407 const MCSectionMachO &Section = cast<MCSectionMachO>(SD.getSection());
408 WriteBytes(Section.getSectionName(), 16);
409 WriteBytes(Section.getSegmentName(), 16);
411 Write64(getSectionAddress(&SD)); // address
412 Write64(SectionSize); // size
414 Write32(getSectionAddress(&SD)); // address
415 Write32(SectionSize); // size
419 unsigned Flags = Section.getTypeAndAttributes();
420 if (SD.hasInstructions())
421 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
423 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
424 Write32(Log2_32(SD.getAlignment()));
425 Write32(NumRelocations ? RelocationsStart : 0);
426 Write32(NumRelocations);
428 Write32(IndirectSymBase.lookup(&SD)); // reserved1
429 Write32(Section.getStubSize()); // reserved2
431 Write32(0); // reserved3
433 assert(OS.tell() - Start == (is64Bit() ? macho::Section64Size :
434 macho::Section32Size));
437 void MachObjectWriter::WriteSymtabLoadCommand(uint32_t SymbolOffset,
439 uint32_t StringTableOffset,
440 uint32_t StringTableSize) {
441 // struct symtab_command (24 bytes)
443 uint64_t Start = OS.tell();
446 Write32(macho::LCT_Symtab);
447 Write32(macho::SymtabLoadCommandSize);
448 Write32(SymbolOffset);
450 Write32(StringTableOffset);
451 Write32(StringTableSize);
453 assert(OS.tell() - Start == macho::SymtabLoadCommandSize);
456 void MachObjectWriter::WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
457 uint32_t NumLocalSymbols,
458 uint32_t FirstExternalSymbol,
459 uint32_t NumExternalSymbols,
460 uint32_t FirstUndefinedSymbol,
461 uint32_t NumUndefinedSymbols,
462 uint32_t IndirectSymbolOffset,
463 uint32_t NumIndirectSymbols) {
464 // struct dysymtab_command (80 bytes)
466 uint64_t Start = OS.tell();
469 Write32(macho::LCT_Dysymtab);
470 Write32(macho::DysymtabLoadCommandSize);
471 Write32(FirstLocalSymbol);
472 Write32(NumLocalSymbols);
473 Write32(FirstExternalSymbol);
474 Write32(NumExternalSymbols);
475 Write32(FirstUndefinedSymbol);
476 Write32(NumUndefinedSymbols);
477 Write32(0); // tocoff
479 Write32(0); // modtaboff
480 Write32(0); // nmodtab
481 Write32(0); // extrefsymoff
482 Write32(0); // nextrefsyms
483 Write32(IndirectSymbolOffset);
484 Write32(NumIndirectSymbols);
485 Write32(0); // extreloff
486 Write32(0); // nextrel
487 Write32(0); // locreloff
488 Write32(0); // nlocrel
490 assert(OS.tell() - Start == macho::DysymtabLoadCommandSize);
493 void MachObjectWriter::WriteNlist(MachSymbolData &MSD,
494 const MCAsmLayout &Layout) {
495 MCSymbolData &Data = *MSD.SymbolData;
496 const MCSymbol &Symbol = Data.getSymbol();
498 uint16_t Flags = Data.getFlags();
499 uint32_t Address = 0;
501 // Set the N_TYPE bits. See <mach-o/nlist.h>.
503 // FIXME: Are the prebound or indirect fields possible here?
504 if (Symbol.isUndefined())
505 Type = macho::STT_Undefined;
506 else if (Symbol.isAbsolute())
507 Type = macho::STT_Absolute;
509 Type = macho::STT_Section;
511 // FIXME: Set STAB bits.
513 if (Data.isPrivateExtern())
514 Type |= macho::STF_PrivateExtern;
517 if (Data.isExternal() || Symbol.isUndefined())
518 Type |= macho::STF_External;
520 // Compute the symbol address.
521 if (Symbol.isDefined()) {
522 if (Symbol.isAbsolute()) {
523 Address = cast<MCConstantExpr>(Symbol.getVariableValue())->getValue();
525 Address = getSymbolAddress(&Data, Layout);
527 } else if (Data.isCommon()) {
528 // Common symbols are encoded with the size in the address
529 // field, and their alignment in the flags.
530 Address = Data.getCommonSize();
532 // Common alignment is packed into the 'desc' bits.
533 if (unsigned Align = Data.getCommonAlignment()) {
534 unsigned Log2Size = Log2_32(Align);
535 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
537 report_fatal_error("invalid 'common' alignment '" +
539 // FIXME: Keep this mask with the SymbolFlags enumeration.
540 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
544 // struct nlist (12 bytes)
546 Write32(MSD.StringIndex);
548 Write8(MSD.SectionIndex);
550 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
559 void MachObjectWriter::RecordX86_64Relocation(const MCAssembler &Asm,
560 const MCAsmLayout &Layout,
561 const MCFragment *Fragment,
562 const MCFixup &Fixup,
564 uint64_t &FixedValue) {
565 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
566 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
567 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
570 uint32_t FixupOffset =
571 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
572 uint32_t FixupAddress =
573 getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
576 unsigned IsExtern = 0;
579 Value = Target.getConstant();
582 // Compensate for the relocation offset, Darwin x86_64 relocations only have
583 // the addend and appear to have attempted to define it to be the actual
584 // expression addend without the PCrel bias. However, instructions with data
585 // following the relocation are not accommodated for (see comment below
586 // regarding SIGNED{1,2,4}), so it isn't exactly that either.
587 Value += 1LL << Log2Size;
590 if (Target.isAbsolute()) { // constant
591 // SymbolNum of 0 indicates the absolute section.
592 Type = macho::RIT_X86_64_Unsigned;
595 // FIXME: I believe this is broken, I don't think the linker can understand
596 // it. I think it would require a local relocation, but I'm not sure if that
597 // would work either. The official way to get an absolute PCrel relocation
598 // is to use an absolute symbol (which we don't support yet).
601 Type = macho::RIT_X86_64_Branch;
603 } else if (Target.getSymB()) { // A - B + constant
604 const MCSymbol *A = &Target.getSymA()->getSymbol();
605 MCSymbolData &A_SD = Asm.getSymbolData(*A);
606 const MCSymbolData *A_Base = Asm.getAtom(&A_SD);
608 const MCSymbol *B = &Target.getSymB()->getSymbol();
609 MCSymbolData &B_SD = Asm.getSymbolData(*B);
610 const MCSymbolData *B_Base = Asm.getAtom(&B_SD);
612 // Neither symbol can be modified.
613 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
614 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
615 report_fatal_error("unsupported relocation of modified symbol");
617 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
618 // implement most of these correctly.
620 report_fatal_error("unsupported pc-relative relocation of difference");
622 // The support for the situation where one or both of the symbols would
623 // require a local relocation is handled just like if the symbols were
624 // external. This is certainly used in the case of debug sections where the
625 // section has only temporary symbols and thus the symbols don't have base
626 // symbols. This is encoded using the section ordinal and non-extern
627 // relocation entries.
629 // Darwin 'as' doesn't emit correct relocations for this (it ends up with a
630 // single SIGNED relocation); reject it for now. Except the case where both
631 // symbols don't have a base, equal but both NULL.
632 if (A_Base == B_Base && A_Base)
633 report_fatal_error("unsupported relocation with identical base");
635 Value += getSymbolAddress(&A_SD, Layout) -
636 (A_Base == NULL ? 0 : getSymbolAddress(A_Base, Layout));
637 Value -= getSymbolAddress(&B_SD, Layout) -
638 (B_Base == NULL ? 0 : getSymbolAddress(B_Base, Layout));
641 Index = A_Base->getIndex();
645 Index = A_SD.getFragment()->getParent()->getOrdinal() + 1;
648 Type = macho::RIT_X86_64_Unsigned;
650 macho::RelocationEntry MRE;
651 MRE.Word0 = FixupOffset;
652 MRE.Word1 = ((Index << 0) |
657 Relocations[Fragment->getParent()].push_back(MRE);
660 Index = B_Base->getIndex();
664 Index = B_SD.getFragment()->getParent()->getOrdinal() + 1;
667 Type = macho::RIT_X86_64_Subtractor;
669 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
670 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
671 const MCSymbolData *Base = Asm.getAtom(&SD);
673 // Relocations inside debug sections always use local relocations when
674 // possible. This seems to be done because the debugger doesn't fully
675 // understand x86_64 relocation entries, and expects to find values that
676 // have already been fixed up.
677 if (Symbol->isInSection()) {
678 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
679 Fragment->getParent()->getSection());
680 if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG))
684 // x86_64 almost always uses external relocations, except when there is no
685 // symbol to use as a base address (a local symbol with no preceding
686 // non-local symbol).
688 Index = Base->getIndex();
691 // Add the local offset, if needed.
693 Value += Layout.getSymbolOffset(&SD) - Layout.getSymbolOffset(Base);
694 } else if (Symbol->isInSection() && !Symbol->isVariable()) {
695 // The index is the section ordinal (1-based).
696 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
698 Value += getSymbolAddress(&SD, Layout);
701 Value -= FixupAddress + (1 << Log2Size);
702 } else if (Symbol->isVariable()) {
703 const MCExpr *Value = Symbol->getVariableValue();
705 bool isAbs = Value->EvaluateAsAbsolute(Res, Layout, SectionAddress);
710 report_fatal_error("unsupported relocation of variable '" +
711 Symbol->getName() + "'");
714 report_fatal_error("unsupported relocation of undefined symbol '" +
715 Symbol->getName() + "'");
718 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
721 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
722 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
723 // rewrite the movq to an leaq at link time if the symbol ends up in
724 // the same linkage unit.
725 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
726 Type = macho::RIT_X86_64_GOTLoad;
728 Type = macho::RIT_X86_64_GOT;
729 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
730 Type = macho::RIT_X86_64_TLV;
731 } else if (Modifier != MCSymbolRefExpr::VK_None) {
732 report_fatal_error("unsupported symbol modifier in relocation");
734 Type = macho::RIT_X86_64_Signed;
736 // The Darwin x86_64 relocation format has a problem where it cannot
737 // encode an address (L<foo> + <constant>) which is outside the atom
738 // containing L<foo>. Generally, this shouldn't occur but it does
739 // happen when we have a RIPrel instruction with data following the
740 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
741 // adjustment Darwin x86_64 uses, the offset is still negative and the
742 // linker has no way to recognize this.
744 // To work around this, Darwin uses several special relocation types
745 // to indicate the offsets. However, the specification or
746 // implementation of these seems to also be incomplete; they should
747 // adjust the addend as well based on the actual encoded instruction
748 // (the additional bias), but instead appear to just look at the final
750 switch (-(Target.getConstant() + (1LL << Log2Size))) {
751 case 1: Type = macho::RIT_X86_64_Signed1; break;
752 case 2: Type = macho::RIT_X86_64_Signed2; break;
753 case 4: Type = macho::RIT_X86_64_Signed4; break;
757 if (Modifier != MCSymbolRefExpr::VK_None)
758 report_fatal_error("unsupported symbol modifier in branch "
761 Type = macho::RIT_X86_64_Branch;
764 if (Modifier == MCSymbolRefExpr::VK_GOT) {
765 Type = macho::RIT_X86_64_GOT;
766 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
767 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in which
768 // case all we do is set the PCrel bit in the relocation entry; this is
769 // used with exception handling, for example. The source is required to
770 // include any necessary offset directly.
771 Type = macho::RIT_X86_64_GOT;
773 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
774 report_fatal_error("TLVP symbol modifier should have been rip-rel");
775 } else if (Modifier != MCSymbolRefExpr::VK_None)
776 report_fatal_error("unsupported symbol modifier in relocation");
778 Type = macho::RIT_X86_64_Unsigned;
782 // x86_64 always writes custom values into the fixups.
785 // struct relocation_info (8 bytes)
786 macho::RelocationEntry MRE;
787 MRE.Word0 = FixupOffset;
788 MRE.Word1 = ((Index << 0) |
793 Relocations[Fragment->getParent()].push_back(MRE);
796 void MachObjectWriter::RecordScatteredRelocation(const MCAssembler &Asm,
797 const MCAsmLayout &Layout,
798 const MCFragment *Fragment,
799 const MCFixup &Fixup,
802 uint64_t &FixedValue) {
803 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
804 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
805 unsigned Type = macho::RIT_Vanilla;
808 const MCSymbol *A = &Target.getSymA()->getSymbol();
809 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
811 if (!A_SD->getFragment())
812 report_fatal_error("symbol '" + A->getName() +
813 "' can not be undefined in a subtraction expression");
815 uint32_t Value = getSymbolAddress(A_SD, Layout);
816 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
817 FixedValue += SecAddr;
820 if (const MCSymbolRefExpr *B = Target.getSymB()) {
821 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
823 if (!B_SD->getFragment())
824 report_fatal_error("symbol '" + B->getSymbol().getName() +
825 "' can not be undefined in a subtraction expression");
827 // Select the appropriate difference relocation type.
829 // Note that there is no longer any semantic difference between these two
830 // relocation types from the linkers point of view, this is done solely for
831 // pedantic compatibility with 'as'.
832 Type = A_SD->isExternal() ? (unsigned)macho::RIT_Difference :
833 (unsigned)macho::RIT_Generic_LocalDifference;
834 Value2 = getSymbolAddress(B_SD, Layout);
835 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
838 // Relocations are written out in reverse order, so the PAIR comes first.
839 if (Type == macho::RIT_Difference ||
840 Type == macho::RIT_Generic_LocalDifference) {
841 macho::RelocationEntry MRE;
842 MRE.Word0 = ((0 << 0) |
843 (macho::RIT_Pair << 24) |
846 macho::RF_Scattered);
848 Relocations[Fragment->getParent()].push_back(MRE);
851 macho::RelocationEntry MRE;
852 MRE.Word0 = ((FixupOffset << 0) |
856 macho::RF_Scattered);
858 Relocations[Fragment->getParent()].push_back(MRE);
861 void MachObjectWriter::RecordARMScatteredRelocation(const MCAssembler &Asm,
862 const MCAsmLayout &Layout,
863 const MCFragment *Fragment,
864 const MCFixup &Fixup,
867 uint64_t &FixedValue) {
868 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
869 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
870 unsigned Type = macho::RIT_Vanilla;
873 const MCSymbol *A = &Target.getSymA()->getSymbol();
874 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
876 if (!A_SD->getFragment())
877 report_fatal_error("symbol '" + A->getName() +
878 "' can not be undefined in a subtraction expression");
880 uint32_t Value = getSymbolAddress(A_SD, Layout);
881 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
882 FixedValue += SecAddr;
885 if (const MCSymbolRefExpr *B = Target.getSymB()) {
886 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
888 if (!B_SD->getFragment())
889 report_fatal_error("symbol '" + B->getSymbol().getName() +
890 "' can not be undefined in a subtraction expression");
892 // Select the appropriate difference relocation type.
893 Type = macho::RIT_Difference;
894 Value2 = getSymbolAddress(B_SD, Layout);
895 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
898 // Relocations are written out in reverse order, so the PAIR comes first.
899 if (Type == macho::RIT_Difference ||
900 Type == macho::RIT_Generic_LocalDifference) {
901 macho::RelocationEntry MRE;
902 MRE.Word0 = ((0 << 0) |
903 (macho::RIT_Pair << 24) |
906 macho::RF_Scattered);
908 Relocations[Fragment->getParent()].push_back(MRE);
911 macho::RelocationEntry MRE;
912 MRE.Word0 = ((FixupOffset << 0) |
916 macho::RF_Scattered);
918 Relocations[Fragment->getParent()].push_back(MRE);
921 void MachObjectWriter::RecordARMMovwMovtRelocation(const MCAssembler &Asm,
922 const MCAsmLayout &Layout,
923 const MCFragment *Fragment,
924 const MCFixup &Fixup,
926 uint64_t &FixedValue) {
927 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
928 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
929 unsigned Type = macho::RIT_ARM_Half;
932 const MCSymbol *A = &Target.getSymA()->getSymbol();
933 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
935 if (!A_SD->getFragment())
936 report_fatal_error("symbol '" + A->getName() +
937 "' can not be undefined in a subtraction expression");
939 uint32_t Value = getSymbolAddress(A_SD, Layout);
941 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
942 FixedValue += SecAddr;
944 if (const MCSymbolRefExpr *B = Target.getSymB()) {
945 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
947 if (!B_SD->getFragment())
948 report_fatal_error("symbol '" + B->getSymbol().getName() +
949 "' can not be undefined in a subtraction expression");
951 // Select the appropriate difference relocation type.
952 Type = macho::RIT_ARM_HalfDifference;
953 Value2 = getSymbolAddress(B_SD, Layout);
954 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
957 // Relocations are written out in reverse order, so the PAIR comes first.
958 // ARM_RELOC_HALF and ARM_RELOC_HALF_SECTDIFF abuse the r_length field:
960 // For these two r_type relocations they always have a pair following them and
961 // the r_length bits are used differently. The encoding of the r_length is as
963 // low bit of r_length:
964 // 0 - :lower16: for movw instructions
965 // 1 - :upper16: for movt instructions
966 // high bit of r_length:
967 // 0 - arm instructions
968 // 1 - thumb instructions
969 // the other half of the relocated expression is in the following pair
970 // relocation entry in the the low 16 bits of r_address field.
971 unsigned ThumbBit = 0;
972 unsigned MovtBit = 0;
973 switch ((unsigned)Fixup.getKind()) {
975 case ARM::fixup_arm_movt_hi16:
976 case ARM::fixup_arm_movt_hi16_pcrel:
979 case ARM::fixup_t2_movt_hi16:
980 case ARM::fixup_t2_movt_hi16_pcrel:
983 case ARM::fixup_t2_movw_lo16:
984 case ARM::fixup_t2_movw_lo16_pcrel:
990 if (Type == macho::RIT_ARM_HalfDifference) {
991 uint32_t OtherHalf = MovtBit
992 ? (FixedValue & 0xffff) : ((FixedValue & 0xffff0000) >> 16);
994 macho::RelocationEntry MRE;
995 MRE.Word0 = ((OtherHalf << 0) |
996 (macho::RIT_Pair << 24) |
1000 macho::RF_Scattered);
1002 Relocations[Fragment->getParent()].push_back(MRE);
1005 macho::RelocationEntry MRE;
1006 MRE.Word0 = ((FixupOffset << 0) |
1011 macho::RF_Scattered);
1013 Relocations[Fragment->getParent()].push_back(MRE);
1016 void MachObjectWriter::RecordTLVPRelocation(const MCAssembler &Asm,
1017 const MCAsmLayout &Layout,
1018 const MCFragment *Fragment,
1019 const MCFixup &Fixup,
1021 uint64_t &FixedValue) {
1022 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
1024 "Should only be called with a 32-bit TLVP relocation!");
1026 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
1027 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
1028 unsigned IsPCRel = 0;
1030 // Get the symbol data.
1031 MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
1032 unsigned Index = SD_A->getIndex();
1034 // We're only going to have a second symbol in pic mode and it'll be a
1035 // subtraction from the picbase. For 32-bit pic the addend is the difference
1036 // between the picbase and the next address. For 32-bit static the addend is
1038 if (Target.getSymB()) {
1039 // If this is a subtraction then we're pcrel.
1040 uint32_t FixupAddress =
1041 getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
1042 MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol());
1044 FixedValue = (FixupAddress - getSymbolAddress(SD_B, Layout) +
1045 Target.getConstant());
1046 FixedValue += 1ULL << Log2Size;
1051 // struct relocation_info (8 bytes)
1052 macho::RelocationEntry MRE;
1054 MRE.Word1 = ((Index << 0) |
1057 (1 << 27) | // Extern
1058 (macho::RIT_Generic_TLV << 28)); // Type
1059 Relocations[Fragment->getParent()].push_back(MRE);
1062 bool MachObjectWriter::getARMFixupKindMachOInfo(unsigned Kind,
1063 unsigned &RelocType,
1064 unsigned &Log2Size) {
1065 RelocType = unsigned(macho::RIT_Vanilla);
1073 Log2Size = llvm::Log2_32(1);
1076 Log2Size = llvm::Log2_32(2);
1079 Log2Size = llvm::Log2_32(4);
1082 Log2Size = llvm::Log2_32(8);
1085 // Handle 24-bit branch kinds.
1086 case ARM::fixup_arm_ldst_pcrel_12:
1087 case ARM::fixup_arm_pcrel_10:
1088 case ARM::fixup_arm_adr_pcrel_12:
1089 case ARM::fixup_arm_condbranch:
1090 case ARM::fixup_arm_uncondbranch:
1091 RelocType = unsigned(macho::RIT_ARM_Branch24Bit);
1092 // Report as 'long', even though that is not quite accurate.
1093 Log2Size = llvm::Log2_32(4);
1096 // Handle Thumb branches.
1097 case ARM::fixup_arm_thumb_br:
1098 RelocType = unsigned(macho::RIT_ARM_ThumbBranch22Bit);
1099 Log2Size = llvm::Log2_32(2);
1102 case ARM::fixup_arm_thumb_bl:
1103 case ARM::fixup_arm_thumb_blx:
1104 RelocType = unsigned(macho::RIT_ARM_ThumbBranch22Bit);
1105 Log2Size = llvm::Log2_32(4);
1108 case ARM::fixup_arm_movt_hi16:
1109 case ARM::fixup_arm_movt_hi16_pcrel:
1110 case ARM::fixup_t2_movt_hi16:
1111 case ARM::fixup_t2_movt_hi16_pcrel:
1112 RelocType = unsigned(macho::RIT_ARM_HalfDifference);
1113 // Report as 'long', even though that is not quite accurate.
1114 Log2Size = llvm::Log2_32(4);
1117 case ARM::fixup_arm_movw_lo16:
1118 case ARM::fixup_arm_movw_lo16_pcrel:
1119 case ARM::fixup_t2_movw_lo16:
1120 case ARM::fixup_t2_movw_lo16_pcrel:
1121 RelocType = unsigned(macho::RIT_ARM_Half);
1122 // Report as 'long', even though that is not quite accurate.
1123 Log2Size = llvm::Log2_32(4);
1127 void MachObjectWriter::RecordARMRelocation(const MCAssembler &Asm,
1128 const MCAsmLayout &Layout,
1129 const MCFragment *Fragment,
1130 const MCFixup &Fixup,
1132 uint64_t &FixedValue) {
1133 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
1135 unsigned RelocType = macho::RIT_Vanilla;
1136 if (!getARMFixupKindMachOInfo(Fixup.getKind(), RelocType, Log2Size)) {
1137 report_fatal_error("unknown ARM fixup kind!");
1141 // If this is a difference or a defined symbol plus an offset, then we need a
1142 // scattered relocation entry. Differences always require scattered
1144 if (Target.getSymB()) {
1145 if (RelocType == macho::RIT_ARM_Half ||
1146 RelocType == macho::RIT_ARM_HalfDifference)
1147 return RecordARMMovwMovtRelocation(Asm, Layout, Fragment, Fixup,
1148 Target, FixedValue);
1149 return RecordARMScatteredRelocation(Asm, Layout, Fragment, Fixup,
1150 Target, Log2Size, FixedValue);
1153 // Get the symbol data, if any.
1154 MCSymbolData *SD = 0;
1155 if (Target.getSymA())
1156 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
1158 // FIXME: For other platforms, we need to use scattered relocations for
1159 // internal relocations with offsets. If this is an internal relocation with
1160 // an offset, it also needs a scattered relocation entry.
1162 // Is this right for ARM?
1163 uint32_t Offset = Target.getConstant();
1164 if (IsPCRel && RelocType == macho::RIT_Vanilla)
1165 Offset += 1 << Log2Size;
1166 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
1167 return RecordARMScatteredRelocation(Asm, Layout, Fragment, Fixup, Target,
1168 Log2Size, FixedValue);
1171 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
1173 unsigned IsExtern = 0;
1176 if (Target.isAbsolute()) { // constant
1178 report_fatal_error("FIXME: relocations to absolute targets "
1179 "not yet implemented");
1181 // Resolve constant variables.
1182 if (SD->getSymbol().isVariable()) {
1184 if (SD->getSymbol().getVariableValue()->EvaluateAsAbsolute(
1185 Res, Layout, SectionAddress)) {
1191 // Check whether we need an external or internal relocation.
1192 if (doesSymbolRequireExternRelocation(SD)) {
1194 Index = SD->getIndex();
1196 // For external relocations, make sure to offset the fixup value to
1197 // compensate for the addend of the symbol address, if it was
1198 // undefined. This occurs with weak definitions, for example.
1199 if (!SD->Symbol->isUndefined())
1200 FixedValue -= Layout.getSymbolOffset(SD);
1202 // The index is the section ordinal (1-based).
1203 const MCSectionData &SymSD = Asm.getSectionData(
1204 SD->getSymbol().getSection());
1205 Index = SymSD.getOrdinal() + 1;
1206 FixedValue += getSectionAddress(&SymSD);
1209 FixedValue -= getSectionAddress(Fragment->getParent());
1211 // The type is determined by the fixup kind.
1215 // struct relocation_info (8 bytes)
1216 macho::RelocationEntry MRE;
1217 MRE.Word0 = FixupOffset;
1218 MRE.Word1 = ((Index << 0) |
1223 Relocations[Fragment->getParent()].push_back(MRE);
1226 void MachObjectWriter::RecordRelocation(const MCAssembler &Asm,
1227 const MCAsmLayout &Layout,
1228 const MCFragment *Fragment,
1229 const MCFixup &Fixup,
1231 uint64_t &FixedValue) {
1232 // FIXME: These needs to be factored into the target Mach-O writer.
1234 RecordARMRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1238 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1242 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
1243 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
1245 // If this is a 32-bit TLVP reloc it's handled a bit differently.
1246 if (Target.getSymA() &&
1247 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
1248 RecordTLVPRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1252 // If this is a difference or a defined symbol plus an offset, then we need a
1253 // scattered relocation entry. Differences always require scattered
1255 if (Target.getSymB())
1256 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
1257 Target, Log2Size, FixedValue);
1259 // Get the symbol data, if any.
1260 MCSymbolData *SD = 0;
1261 if (Target.getSymA())
1262 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
1264 // If this is an internal relocation with an offset, it also needs a scattered
1265 // relocation entry.
1266 uint32_t Offset = Target.getConstant();
1268 Offset += 1 << Log2Size;
1269 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
1270 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
1271 Target, Log2Size, FixedValue);
1274 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
1276 unsigned IsExtern = 0;
1279 if (Target.isAbsolute()) { // constant
1280 // SymbolNum of 0 indicates the absolute section.
1282 // FIXME: Currently, these are never generated (see code below). I cannot
1283 // find a case where they are actually emitted.
1284 Type = macho::RIT_Vanilla;
1286 // Resolve constant variables.
1287 if (SD->getSymbol().isVariable()) {
1289 if (SD->getSymbol().getVariableValue()->EvaluateAsAbsolute(
1290 Res, Layout, SectionAddress)) {
1296 // Check whether we need an external or internal relocation.
1297 if (doesSymbolRequireExternRelocation(SD)) {
1299 Index = SD->getIndex();
1300 // For external relocations, make sure to offset the fixup value to
1301 // compensate for the addend of the symbol address, if it was
1302 // undefined. This occurs with weak definitions, for example.
1303 if (!SD->Symbol->isUndefined())
1304 FixedValue -= Layout.getSymbolOffset(SD);
1306 // The index is the section ordinal (1-based).
1307 const MCSectionData &SymSD = Asm.getSectionData(
1308 SD->getSymbol().getSection());
1309 Index = SymSD.getOrdinal() + 1;
1310 FixedValue += getSectionAddress(&SymSD);
1313 FixedValue -= getSectionAddress(Fragment->getParent());
1315 Type = macho::RIT_Vanilla;
1318 // struct relocation_info (8 bytes)
1319 macho::RelocationEntry MRE;
1320 MRE.Word0 = FixupOffset;
1321 MRE.Word1 = ((Index << 0) |
1326 Relocations[Fragment->getParent()].push_back(MRE);
1329 void MachObjectWriter::BindIndirectSymbols(MCAssembler &Asm) {
1330 // This is the point where 'as' creates actual symbols for indirect symbols
1331 // (in the following two passes). It would be easier for us to do this sooner
1332 // when we see the attribute, but that makes getting the order in the symbol
1333 // table much more complicated than it is worth.
1335 // FIXME: Revisit this when the dust settles.
1337 // Bind non lazy symbol pointers first.
1338 unsigned IndirectIndex = 0;
1339 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
1340 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
1341 const MCSectionMachO &Section =
1342 cast<MCSectionMachO>(it->SectionData->getSection());
1344 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
1347 // Initialize the section indirect symbol base, if necessary.
1348 if (!IndirectSymBase.count(it->SectionData))
1349 IndirectSymBase[it->SectionData] = IndirectIndex;
1351 Asm.getOrCreateSymbolData(*it->Symbol);
1354 // Then lazy symbol pointers and symbol stubs.
1356 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
1357 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
1358 const MCSectionMachO &Section =
1359 cast<MCSectionMachO>(it->SectionData->getSection());
1361 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
1362 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
1365 // Initialize the section indirect symbol base, if necessary.
1366 if (!IndirectSymBase.count(it->SectionData))
1367 IndirectSymBase[it->SectionData] = IndirectIndex;
1369 // Set the symbol type to undefined lazy, but only on construction.
1371 // FIXME: Do not hardcode.
1373 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
1375 Entry.setFlags(Entry.getFlags() | 0x0001);
1379 /// ComputeSymbolTable - Compute the symbol table data
1381 /// \param StringTable [out] - The string table data.
1382 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
1384 void MachObjectWriter::
1385 ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
1386 std::vector<MachSymbolData> &LocalSymbolData,
1387 std::vector<MachSymbolData> &ExternalSymbolData,
1388 std::vector<MachSymbolData> &UndefinedSymbolData) {
1389 // Build section lookup table.
1390 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
1392 for (MCAssembler::iterator it = Asm.begin(),
1393 ie = Asm.end(); it != ie; ++it, ++Index)
1394 SectionIndexMap[&it->getSection()] = Index;
1395 assert(Index <= 256 && "Too many sections!");
1397 // Index 0 is always the empty string.
1398 StringMap<uint64_t> StringIndexMap;
1399 StringTable += '\x00';
1401 // Build the symbol arrays and the string table, but only for non-local
1404 // The particular order that we collect the symbols and create the string
1405 // table, then sort the symbols is chosen to match 'as'. Even though it
1406 // doesn't matter for correctness, this is important for letting us diff .o
1408 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1409 ie = Asm.symbol_end(); it != ie; ++it) {
1410 const MCSymbol &Symbol = it->getSymbol();
1412 // Ignore non-linker visible symbols.
1413 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1416 if (!it->isExternal() && !Symbol.isUndefined())
1419 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1421 Entry = StringTable.size();
1422 StringTable += Symbol.getName();
1423 StringTable += '\x00';
1427 MSD.SymbolData = it;
1428 MSD.StringIndex = Entry;
1430 if (Symbol.isUndefined()) {
1431 MSD.SectionIndex = 0;
1432 UndefinedSymbolData.push_back(MSD);
1433 } else if (Symbol.isAbsolute()) {
1434 MSD.SectionIndex = 0;
1435 ExternalSymbolData.push_back(MSD);
1437 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1438 assert(MSD.SectionIndex && "Invalid section index!");
1439 ExternalSymbolData.push_back(MSD);
1443 // Now add the data for local symbols.
1444 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1445 ie = Asm.symbol_end(); it != ie; ++it) {
1446 const MCSymbol &Symbol = it->getSymbol();
1448 // Ignore non-linker visible symbols.
1449 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1452 if (it->isExternal() || Symbol.isUndefined())
1455 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1457 Entry = StringTable.size();
1458 StringTable += Symbol.getName();
1459 StringTable += '\x00';
1463 MSD.SymbolData = it;
1464 MSD.StringIndex = Entry;
1466 if (Symbol.isAbsolute()) {
1467 MSD.SectionIndex = 0;
1468 LocalSymbolData.push_back(MSD);
1470 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1471 assert(MSD.SectionIndex && "Invalid section index!");
1472 LocalSymbolData.push_back(MSD);
1476 // External and undefined symbols are required to be in lexicographic order.
1477 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
1478 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
1480 // Set the symbol indices.
1482 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1483 LocalSymbolData[i].SymbolData->setIndex(Index++);
1484 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1485 ExternalSymbolData[i].SymbolData->setIndex(Index++);
1486 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1487 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
1489 // The string table is padded to a multiple of 4.
1490 while (StringTable.size() % 4)
1491 StringTable += '\x00';
1494 void MachObjectWriter::computeSectionAddresses(const MCAssembler &Asm,
1495 const MCAsmLayout &Layout) {
1496 uint64_t StartAddress = 0;
1497 const SmallVectorImpl<MCSectionData*> &Order = Layout.getSectionOrder();
1498 for (int i = 0, n = Order.size(); i != n ; ++i) {
1499 const MCSectionData *SD = Order[i];
1500 StartAddress = RoundUpToAlignment(StartAddress, SD->getAlignment());
1501 SectionAddress[SD] = StartAddress;
1502 StartAddress += Layout.getSectionAddressSize(SD);
1504 // Explicitly pad the section to match the alignment requirements of the
1505 // following one. This is for 'gas' compatibility, it shouldn't
1506 /// strictly be necessary.
1507 StartAddress += getPaddingSize(SD, Layout);
1511 void MachObjectWriter::ExecutePostLayoutBinding(MCAssembler &Asm,
1512 const MCAsmLayout &Layout) {
1513 computeSectionAddresses(Asm, Layout);
1515 // Create symbol data for any indirect symbols.
1516 BindIndirectSymbols(Asm);
1518 // Compute symbol table information and bind symbol indices.
1519 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
1520 UndefinedSymbolData);
1523 bool MachObjectWriter::
1524 IsSymbolRefDifferenceFullyResolvedImpl(const MCAssembler &Asm,
1525 const MCSymbolData &DataA,
1526 const MCFragment &FB,
1528 bool IsPCRel) const {
1532 // The effective address is
1533 // addr(atom(A)) + offset(A)
1534 // - addr(atom(B)) - offset(B)
1535 // and the offsets are not relocatable, so the fixup is fully resolved when
1536 // addr(atom(A)) - addr(atom(B)) == 0.
1537 const MCSymbolData *A_Base = 0, *B_Base = 0;
1539 const MCSymbol &SA = DataA.getSymbol().AliasedSymbol();
1540 const MCSection &SecA = SA.getSection();
1541 const MCSection &SecB = FB.getParent()->getSection();
1544 // The simple (Darwin, except on x86_64) way of dealing with this was to
1545 // assume that any reference to a temporary symbol *must* be a temporary
1546 // symbol in the same atom, unless the sections differ. Therefore, any PCrel
1547 // relocation to a temporary symbol (in the same section) is fully
1548 // resolved. This also works in conjunction with absolutized .set, which
1549 // requires the compiler to use .set to absolutize the differences between
1550 // symbols which the compiler knows to be assembly time constants, so we
1551 // don't need to worry about considering symbol differences fully resolved.
1553 if (!Asm.getBackend().hasReliableSymbolDifference()) {
1554 if (!SA.isTemporary() || !SA.isInSection() || &SecA != &SecB)
1559 if (!TargetObjectWriter->useAggressiveSymbolFolding())
1563 const MCFragment &FA = *Asm.getSymbolData(SA).getFragment();
1565 A_Base = FA.getAtom();
1569 B_Base = FB.getAtom();
1573 // If the atoms are the same, they are guaranteed to have the same address.
1574 if (A_Base == B_Base)
1577 // Otherwise, we can't prove this is fully resolved.
1581 void MachObjectWriter::WriteObject(MCAssembler &Asm, const MCAsmLayout &Layout) {
1582 unsigned NumSections = Asm.size();
1584 // The section data starts after the header, the segment load command (and
1585 // section headers) and the symbol table.
1586 unsigned NumLoadCommands = 1;
1587 uint64_t LoadCommandsSize = is64Bit() ?
1588 macho::SegmentLoadCommand64Size + NumSections * macho::Section64Size :
1589 macho::SegmentLoadCommand32Size + NumSections * macho::Section32Size;
1591 // Add the symbol table load command sizes, if used.
1592 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
1593 UndefinedSymbolData.size();
1595 NumLoadCommands += 2;
1596 LoadCommandsSize += (macho::SymtabLoadCommandSize +
1597 macho::DysymtabLoadCommandSize);
1600 // Compute the total size of the section data, as well as its file size and vm
1602 uint64_t SectionDataStart = (is64Bit() ? macho::Header64Size :
1603 macho::Header32Size) + LoadCommandsSize;
1604 uint64_t SectionDataSize = 0;
1605 uint64_t SectionDataFileSize = 0;
1606 uint64_t VMSize = 0;
1607 for (MCAssembler::const_iterator it = Asm.begin(),
1608 ie = Asm.end(); it != ie; ++it) {
1609 const MCSectionData &SD = *it;
1610 uint64_t Address = getSectionAddress(&SD);
1611 uint64_t Size = Layout.getSectionAddressSize(&SD);
1612 uint64_t FileSize = Layout.getSectionFileSize(&SD);
1613 FileSize += getPaddingSize(&SD, Layout);
1615 VMSize = std::max(VMSize, Address + Size);
1617 if (SD.getSection().isVirtualSection())
1620 SectionDataSize = std::max(SectionDataSize, Address + Size);
1621 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
1624 // The section data is padded to 4 bytes.
1626 // FIXME: Is this machine dependent?
1627 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
1628 SectionDataFileSize += SectionDataPadding;
1630 // Write the prolog, starting with the header and load command...
1631 WriteHeader(NumLoadCommands, LoadCommandsSize,
1632 Asm.getSubsectionsViaSymbols());
1633 WriteSegmentLoadCommand(NumSections, VMSize,
1634 SectionDataStart, SectionDataSize);
1636 // ... and then the section headers.
1637 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
1638 for (MCAssembler::const_iterator it = Asm.begin(),
1639 ie = Asm.end(); it != ie; ++it) {
1640 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1641 unsigned NumRelocs = Relocs.size();
1642 uint64_t SectionStart = SectionDataStart + getSectionAddress(it);
1643 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1644 RelocTableEnd += NumRelocs * macho::RelocationInfoSize;
1647 // Write the symbol table load command, if used.
1649 unsigned FirstLocalSymbol = 0;
1650 unsigned NumLocalSymbols = LocalSymbolData.size();
1651 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1652 unsigned NumExternalSymbols = ExternalSymbolData.size();
1653 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1654 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1655 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1656 unsigned NumSymTabSymbols =
1657 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1658 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1659 uint64_t IndirectSymbolOffset = 0;
1661 // If used, the indirect symbols are written after the section data.
1662 if (NumIndirectSymbols)
1663 IndirectSymbolOffset = RelocTableEnd;
1665 // The symbol table is written after the indirect symbol data.
1666 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1668 // The string table is written after symbol table.
1669 uint64_t StringTableOffset =
1670 SymbolTableOffset + NumSymTabSymbols * (is64Bit() ? macho::Nlist64Size :
1671 macho::Nlist32Size);
1672 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1673 StringTableOffset, StringTable.size());
1675 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1676 FirstExternalSymbol, NumExternalSymbols,
1677 FirstUndefinedSymbol, NumUndefinedSymbols,
1678 IndirectSymbolOffset, NumIndirectSymbols);
1681 // Write the actual section data.
1682 for (MCAssembler::const_iterator it = Asm.begin(),
1683 ie = Asm.end(); it != ie; ++it) {
1684 Asm.WriteSectionData(it, Layout);
1686 uint64_t Pad = getPaddingSize(it, Layout);
1687 for (unsigned int i = 0; i < Pad; ++i)
1691 // Write the extra padding.
1692 WriteZeros(SectionDataPadding);
1694 // Write the relocation entries.
1695 for (MCAssembler::const_iterator it = Asm.begin(),
1696 ie = Asm.end(); it != ie; ++it) {
1697 // Write the section relocation entries, in reverse order to match 'as'
1698 // (approximately, the exact algorithm is more complicated than this).
1699 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1700 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1701 Write32(Relocs[e - i - 1].Word0);
1702 Write32(Relocs[e - i - 1].Word1);
1706 // Write the symbol table data, if used.
1708 // Write the indirect symbol entries.
1709 for (MCAssembler::const_indirect_symbol_iterator
1710 it = Asm.indirect_symbol_begin(),
1711 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1712 // Indirect symbols in the non lazy symbol pointer section have some
1713 // special handling.
1714 const MCSectionMachO &Section =
1715 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1716 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1717 // If this symbol is defined and internal, mark it as such.
1718 if (it->Symbol->isDefined() &&
1719 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1720 uint32_t Flags = macho::ISF_Local;
1721 if (it->Symbol->isAbsolute())
1722 Flags |= macho::ISF_Absolute;
1728 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1731 // FIXME: Check that offsets match computed ones.
1733 // Write the symbol table entries.
1734 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1735 WriteNlist(LocalSymbolData[i], Layout);
1736 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1737 WriteNlist(ExternalSymbolData[i], Layout);
1738 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1739 WriteNlist(UndefinedSymbolData[i], Layout);
1741 // Write the string table.
1742 OS << StringTable.str();
1746 MCObjectWriter *llvm::createMachObjectWriter(MCMachObjectTargetWriter *MOTW,
1748 bool IsLittleEndian) {
1749 return new MachObjectWriter(MOTW, OS, IsLittleEndian);